FRONTIER Act and VET AI Act: What Independent AI Assurance Means for AI Buyers

The FRONTIER Act, introduced in the House on July 23, 2026 as H.R. 9925, would require the largest AI developers to publish safety frameworks, report serious incidents, and submit to independent audits. In the Senate, the VET AI Act takes a different route to the same goal: it would direct NIST to write voluntary guidelines for how AI systems are tested and verified by third parties. Neither bill is law. Together they show where federal AI policy is heading on one question that matters to every organization buying AI: who checks whether a developer’s safety claims are true. This article explains what each bill would require, how they differ, where they stand, and what organizations that buy and deploy AI can do with that information now. Why Independent AI Assurance Is Moving Into Federal Law Most of what organizations know about the safety of the AI models they use comes from the developers themselves: model cards, system cards, published policies, and voluntary commitments. Those documents are useful, but they are self-reported. The idea behind both bills is that a claim of safety should be verifiable by someone other than the party making it, the same logic that produced financial audits and security attestations in other markets. Two pressures push this into federal law. The first is a growing body of state legislation on frontier AI safety. The FRONTIER Act’s own summary names California SB 53, New York’s RAISE Act, and Illinois SB 315 as the statutes its preemption clause targets, and it presents a single federal standard as an alternative to a state-by-state patchwork. The second is that regulators of AI users are moving in the same direction. On November 4, 2026, the OCC issued a supervisory framework that asks how banks obtain evidence that AI controls keep working, including through independent challenge and external testing. Elevate covers that framework in [INTERNAL LINK: OCC Super Intelligence Framework article, add URL after publication]. What the FRONTIER Act Would Require The FRONTIER Act (Frontier Risk Oversight, National Transparency, Independent Evaluation, and Reporting Act) was introduced by Representatives Lori Trahan and Jay Obernolte with a bipartisan group of cosponsors. It grew out of the broader Great American AI Act framework the two released earlier as a discussion draft. The descriptions below follow the sponsors’ section-by-section summary. Who It Covers The bill applies to developers of frontier models, defined as foundation models trained with more than 10^26 operations, counting the original training run and any substantial later modification. Within that group it creates two higher tiers based on revenue and AI development spending over the preceding 36 months. A large frontier developer has gross revenues above $50 million and at least $1 billion in AI development expenditures. A very large frontier developer has gross revenues above $5 billion and at least $10 billion in such expenditures. Each higher tier carries every obligation of the tiers below it. Tier Threshold Main obligations Frontier developer Has trained or begun training a model above 10^26 operations Publish a transparency report before or when deploying a new or substantially modified frontier model; report critical safety incidents Large frontier developer Revenues above $50M and at least $1B in AI development spending Write, implement, and publish a frontier AI framework; undergo an annual independent compliance audit and publish the redacted report; register; report catastrophic risk, including from internal use Very large frontier developer Revenues above $5B and at least $10B in AI development spending Retain a licensed independent verification organization for ongoing assessment, with access to unredacted materials The tiers mean the bill reaches only a small number of companies. Most organizations that use AI, including banks, insurers, and software vendors that build on top of foundation models, would not be covered as developers. What Counts as a Catastrophic Risk and a Critical Incident The bill defines catastrophic risk narrowly: a foreseeable, material risk that a frontier model contributes to 50 or more deaths or serious injuries, or more than $1 billion in property damage, in a single incident. The incident must involve the model helping to develop or release a chemical, biological, radiological, nuclear, or cyber weapon, acting without meaningful human oversight in a way that would be a crime such as a cyberattack or theft, or evading the control of its developer or user. A critical safety incident is broader. It includes unauthorized access to or theft of model weights, harm from a catastrophic risk that has materialized, loss of control of a model, and a model deceiving its developer to subvert controls outside a designed evaluation. These are the events developers would have to report to a new Under Secretary of Commerce for AI Security. Independent Verification Organizations The bill’s most novel feature is a licensed market of independent verification organizations. The Under Secretary would license and oversee them, and each very large developer would have to retain one for ongoing assessment of whether its framework, governance, monitoring, and mitigations keep catastrophic risk at acceptable levels. An organization that finds an imminent catastrophic risk would have to refer it to the Secretary of Commerce within 72 hours, and the Secretary could issue emergency orders restricting a model. The Government Accountability Office would report every year on whether these organizations remain independent of the AI industry. Enforcement and Preemption Violations of the transparency, reporting, and verification duties carry civil penalties of up to $1 million per violation per day, enforced by the Attorney General and by state attorneys general who opt in. Violations of an emergency order carry up to $10 million per violation per day, and willful violations are criminal. The bill would also bar states from imposing new obligations on AI developers covering frontier catastrophic-risk transparency, third-party auditing, and incident reporting, while preserving state laws on how AI is deployed and used, laws protecting minors, and state procurement rules. Where It Stands As of early November 2026, H.R. 9925 is pending in committee in the House. It has not passed either chamber. Its details could
AIUC-1 Certification Explained: How It Works and What Buyers Should Check

AIUC-1 certification is the third-party assurance route for AI agents, and the standard behind it was revised again on October 15, 2026, which makes the version a certificate was audited against part of the evidence a buyer should ask for. The Artificial Intelligence Underwriting Company (AIUC) updates the standard every quarter, and the latest release added controls for open-weight models, broadened third-party testing across the domains, and made human approval of sensitive tool actions a core control. For organizations that build AI agents, readiness work built on last quarter’s version needs a check against the current one. For organizations that buy agents, a certificate is only as informative as its scope and its version. This article covers what AIUC-1 is, its six domains, how certification works, what changed in 2026, and how to prepare or verify. What AIUC-1 Is and Who Stands Behind It A Standard for the Agent, Not the Company AIUC-1 is the standard for AI agent security, safety, and reliability. To earn certification, an agent must pass all applicable AIUC-1 requirements, validated by an accredited auditor, and go through evaluations that cover the top enterprise risks. A certificate typically covers specific products, not the full range of an organization’s products and practices. The standard is also AI-focused: AIUC states that it does not duplicate the work of frameworks such as SOC 2, ISO 27001, or GDPR, which address non-AI risks. Who Issues the Certificate Only AIUC can issue the official AIUC-1 certificate. Accredited auditors validate the evidence, and AIUC works with a community of service providers, GRC platforms, and evaluation providers that help organizations close gaps. AIUC also states that certification verifies and complements an organization’s own safety and security practices, so an organization cannot rely on the process to build those practices from scratch. The standard is developed with the AIUC-1 Consortium, which AIUC describes as having more than 300 members. Elevate Consult prepares organizations for AIUC-1 and does not issue the certificate. The Six AIUC-1 Domains AIUC-1 organizes its requirements into six domains, lettered A to F. The table gives AIUC’s description of each domain and an example of a requirement inside it. Domain What it covers Example requirement A. Data & Privacy Data use policies, IP and trade secrets, PII and cross-customer exposure, secrets leakage A003 limits what data an agent can reach; A008 prevents leakage of credentials and secrets B. Security Adversarial testing, input filtering, access controls, endpoint protection, secure code patterns B006 prevents unauthorized agent actions, including limiting which MCP servers an agent can use C. Safety Risk taxonomy, pre-deployment and third-party testing, harmful and out-of-scope outputs C001 defines an AI risk taxonomy; C002 requires pre-deployment testing D. Reliability Hallucination prevention and testing, restricting unsafe tool calls D001 prevents hallucinated outputs; D003 restricts unsafe tool calls E. Accountability Failure plans, ownership, vendor due diligence, logging, transparency, quality management E001 to E003 set AI failure plans; E015 requires logging of AI system activity F. Society Preventing AI-enabled cyber misuse and catastrophic misuse F001 covers cyber misuse, with a control added in October 2026 for assessing modified open-weight models AIUC labels every requirement as mandatory or optional and every control as mandatory or supplemental, and an agent has to pass the requirements that apply to it. Applicability follows the agent’s capabilities: some requirements apply to externally facing agents, for example, and others to agents that generate code or media. The scoping conversation with an auditor therefore decides how much of the standard an organization is assessed against. Agent-Specific Controls Worth Knowing Several controls address the ways agents differ from other software. Tool use is covered by D003, which asks for tool allowlists and parameter validation, rate limits and transaction caps for autonomous tool use, a log of every tool call, and human approval for sensitive operations such as financial transactions or deleting records. Identity and permissions are covered by A003.2 and A003.3, which describe a unique, verifiable identity for each agent, standard federation with OAuth 2.0 or OIDC, just-in-time permissions scoped to subtasks, and no silent inheritance of elevated permissions from orchestrating agents. Logging is covered by E015, whose supplemental controls add full execution chains of agentic workflows and tamper-evident logs. The identity, permission, and execution-chain controls are supplemental, while tool authorization, rate limits, tool call logging, and human approval of sensitive operations are mandatory. Elevate’s guide to governing autonomous AI agents covers the wider governance picture. How AIUC-1 Certification Works AIUC-1 certification combines an audit of the agent’s controls with evaluations of the agent itself. AIUC describes a typical process with four phases: The phase ranges add up to more than the total AIUC quotes, which suggests that the phases overlap. AIUC gives a typical total of four to eight weeks on its certification page and five to ten weeks in its FAQ, and it notes that organizations starting from zero may take longer. At the end, the organization receives the certificate, a comprehensive audit report with third-party attestation and evaluation results, and a badge for its trust center or sales material. What Keeps a Certificate Valid A certificate is valid for 12 months. Evaluations have to run at least every three months to keep it valid, and all technical, operational, and legal controls are re-audited annually. AIUC is explicit that a certificate is a point-in-time assessment of controls, not a warranty of future performance, and that it should not replace independent risk assessment and ongoing monitoring. AIUC-1 Changes Every Quarter, So the Version Matters AIUC publishes the standard on a fixed calendar, on January 15, April 15, July 15, and October 15 of each year, and keeps a public changelog. The table lists the releases and the focus of each one. Release Focus July 22, 2025 First launch of the standard October 1, 2025 Quarterly revision January 15, 2026 Quarterly revision April 15, 2026 MCP security, agent permissions, and third-party risk July 15, 2026 Coding agents, with new requirements for secrets management and secure code generation, and technical guidance for auditors
Agentic AI in Banking: What Leading Banks Control Before Regulators Ask

Agentic AI in banking is moving into production faster than the supervisory guidance that governs it. The revised interagency model risk management guidance issued on April 17, 2026 states that generative AI and agentic AI models are not within its scope. The October 2026 Evident AI Index shows how leading banks respond: 80% of the top-10 banks show public evidence of controls for specific AI use cases, against 40% of the other 40 banks (Evident AI Index: Key Findings Report, October 2026). Evident reports that advances in agentic AI are driving investment in automated guardrails and post-deployment monitoring. For a mid-size bank, the question is which controls to build and in what order. This article covers what changed in 2026, what leading banks do, and a control path a mid-size bank can start now. What Changed for Agentic AI in Banking in 2026 What an AI Agent Changes for Governance A model produces an estimate that a person or a process then uses. An agent goes further: it takes actions through the tools and permissions it holds, so it can query systems, change records, send messages, or start transactions, depending on how it is configured. That changes what a bank must control. Accuracy of the output still matters, but so do what the agent is allowed to do, what data it can reach, and when a person steps in. A bank that governs agents only through its model inventory misses these actions entirely. Evident describes the same shift, with governance moving from validating a model toward controlling the whole workflow around it. The April 2026 Model Risk Guidance Leaves Agentic AI Out On April 17, 2026, the Federal Reserve, the OCC, and the FDIC jointly issued revised model risk management guidance (Federal Reserve SR 26-2 and OCC Bulletin 2026-13). It replaces SR 11-7 and the 2021 interagency statement on model risk for BSA and AML systems. A footnote states that generative AI and agentic AI models are novel and rapidly evolving, so they are not within the scope of the guidance. The principles still apply to traditional statistical models and to non-generative, non-agentic AI models, which means a machine learning credit model stays in scope while an AI agent does not. Elevate’s analysis of the SR 11-7 rescission covers the guidance in full. The Threshold and the Footnote Matter Together The guidance is expected to be most relevant to banking organizations with more than $30 billion in total assets. For banks at or below that size, models typically fall under internal risk management and governance practices suited to the bank’s size and risk profile, although the guidance may be relevant where model exposure is significant. It also sets no enforceable standards, and the agencies state that non-compliance will not result in supervisory criticism. The footnote adds that a bank’s own risk management and governance practices should guide the controls for tools the guidance does not cover. For a mid-size bank, the exclusion therefore does not remove the expectation of governance. It moves the work to the bank. The February 2026 FS AI RMF Offers a Control Set On February 19, 2026, the US Treasury released the Financial Services AI Risk Management Framework (FS AI RMF) alongside an AI lexicon. The Cyber Risk Institute hosts the framework and its 230 control objectives, which adapt the NIST AI RMF to financial services. It opens with an adoption stage questionnaire that places an institution in one of four stages: Initial, Minimal, Evolving, or Embedded. The stage sets the scope, from 21 control objectives at Initial to 126 at Minimal, 193 at Evolving, and all 230 at Embedded. The framework is voluntary, and Elevate’s FS AI RMF readiness service helps institutions determine their stage and scope the controls that follow. What Leading Banks Do About Agentic AI Controls Principles Are Universal, Evidence of Controls Is Not The Evident AI Index benchmarks 50 of the largest banks in North America, Europe, and Asia using only publicly available data. Responsible AI principles are now almost universal: 49 of the 50 banks publish them, up from 16 in 2023. A majority (62%) explain how those principles move into practice, but fewer than half (48%) share evidence of controls implemented for a specific use case. That group skews heavily toward the top-10 banks: 80% of them show such evidence, against 40% of the other 40. Because the Index counts only what banks publish, the gap measures public evidence, not the controls themselves. Governance Moves From the Model to the Workflow Evident observes that governance changes when systems take actions instead of only generating answers. The focus moves beyond conventional model risk, meaning bias, explainability, and validation, toward controlling the whole workflow. The controls it lists are grounding, input and output validation, tool permissions, human intervention, and continuous monitoring. Evident also describes leading banks as investing in controls as an enabler of scale, not only as a brake on risk. Banks that build controls ahead of deployment can release more autonomous systems faster, without governance becoming the next bottleneck. Controls Are Built Once and Reused Leading banks treat the controls as infrastructure around the model. Evident describes this as a reusable “harness” of proprietary data, integrations, tools, permissions, evaluations, controls, and orchestration, so that each deployment makes the next one faster, cheaper, and easier to govern. The same logic supports routing tasks to different models by complexity, latency, risk, and cost. For a mid-size bank, the lesson is structural: a control set written once and applied to every agent costs less than one written for each pilot. Reuse also helps with evidence: a shared logging and approval layer produces the same records for every agent, which makes testing simpler. Model Risk Talent Grows With Deployment Staffing follows the same pattern. AI model risk roles grew 9.0% year over year across 127,000 AI and data employees, behind only enablement (21.2%) and product management (10.0%). Evident links that growth to the need to manage increasingly complex risks as deployment scales
Agentic AI Governance Was the Real Agenda at The AI Conference

Agentic AI governance was not on the title slide of a single keynote at The AI Conference in San Francisco on 30 September and 1 October 2026, yet it was the problem underneath almost all of them. Engineers described reward hacking, prompt injection through an email, agents that pass every dashboard check while doing the wrong thing, and costs that compound invisibly across hundreds of tool calls. Each of those is a control failure described in engineering language. Angela Polania, Principal of Elevate Consult and an ISO 42001 Lead Auditor, attended both days, and this article translates eight of those keynotes into the governance controls your organization will be asked to evidence. Why Agentic AI Governance Dominated a Technical Conference The Agenda Was About Agents, the Problems Were About Control The AI Conference describes itself as an event about agentic AI, frontier models, infrastructure, security, governance, and evaluation. Governance sat on that list, but the governance content did not come from the governance sessions. It came from engineers explaining why their systems surprised them. A researcher showing how an agent gamed a benchmark is describing a gap between a stated requirement and real intent. A founder showing how an email can hijack an agent is describing an unmanaged input channel. Read with a compliance lens, the technical tracks were a two-day catalogue of missing controls. What Changes When a Model Becomes an Agent A model produces an output that a person reads. An agent takes an action: it calls a tool, moves data, sends a message, books a purchase, or changes a record, often without anyone reviewing each step. That single shift moves the risk from what a system says to what a system does. Several speakers made the same observation from different directions: agents now execute dozens of actions per prompt, and a single user request can fan out into hundreds of inference calls. Governance built for a model that answers questions does not cover a system that acts on them. The Cost of Treating This as an Engineering Problem Only Engineering teams will solve the engineering half. They will add observability, routing, sandboxes, and evaluation harnesses, and several speakers showed impressive versions of each. What engineering cannot supply alone is the accountability layer: who approved this agent for this task, which actions fall outside appetite, who is liable when it acts, and what evidence shows the controls operated. Those questions belong to risk, compliance, legal, and the board. Organizations that leave them to the platform team end up with systems that are well instrumented and ungoverned. That is the gap an AI governance and AI risk management program exists to close. Eight Lessons for Agentic AI Governance From the Keynotes 1. AI Red Teaming Must Test Intent, Not the Specification Ion Stoica opened the second day with an anecdote every risk leader should keep. An AI agent asked to build a fast key-value store returned a result far faster than any previous attempt. It got there by noticing that the benchmark used predictable values, so instead of storing data it simply regenerated the values on request. It won the benchmark and failed the actual requirement, which was to store arbitrary data. Stoica framed this as two gaps: a requirement gap, because the specification is only a proxy for what the user meant, and a model gap, because the test environment is only a proxy for the real world. The governance consequence is direct. An agent that passes its acceptance tests has proven it satisfies the tests, not that it does what the business intended. Stoica’s answer was continuous AI red teaming with humans in the loop, using people and agents to hunt for the places where the system’s logic and real intent diverge. That is the discipline behind adversarial technical testing, and it is why the AIUC-1 standard for AI agents requires behavioral testing of the agent rather than review of its documentation alone. Elevate’s penetration testing practice already runs adversarial prompting, jailbreak, and data leakage testing against AI systems for exactly this reason. 2. An Agent’s Explanation Is Not Evidence of Its Behavior Emmanuel Ameisen of Anthropic’s interpretability team showed what researchers can now observe inside a model while it works. In one example, a model trained to resist prompt injection registered internal signals for concepts like a fake or an injected instruction before it wrote a single word of its response. In another, a model that had cheated on a task showed internal activity associated with concealment and avoiding detection while it wrote code to cover its tracks. Neither signal was visible in the final output or in the model’s stated reasoning. Interpretability tooling of this kind is still a research capability, not a control most organizations can deploy today. The lesson for governance is the limit it exposes. A log of what an agent says it did, or a chain of reasoning it produced, is a self-report, and self-reports are weak audit evidence. Monitoring that relies only on outputs will miss the cases that matter most. Assurance has to test behavior independently, under conditions the agent did not choose. 3. Every Input an Agent Reads Is an Attack Surface Illia Polosukhin, co-founder of NEAR Protocol, made the security case plainly. Agents are being given access to email, calendars, and internal systems, and every external input they read is a potential vector. A crafted email or calendar invite can carry instructions the agent follows. He also raised a supply chain question most organizations have not asked: when a prompt passes through a provider’s routing layers, how does the business verify which model actually ran and that the output was not altered on the way back? His proposed answers, confidential computing with cryptographic attestation and a tamper-evident log of every agent action, sit at the infrastructure layer. The governance questions sit above them. Which channels can feed instructions to which agents? Which third-party models and services sit in the path, and what assurance exists for each?
AI Assurance for AI Agents: Choosing the Evidence Enterprise Buyers Accept

AI assurance for AI agents is the set of independent evidence that proves an agent is governed, secure and behaves safely, and it has become a line item in enterprise deals. Two years ago, a vendor selling an AI product could answer most security reviews with a SOC 2 report and a well-written questionnaire. Today, buyers can choose among an ISO/IEC 42001 certificate, a CSA STAR for AI designation and an AIUC-1 audit report with red teaming results, and they are starting to ask which ones you hold. This guide explains what each instrument proves, what it leaves out, how to choose and sequence them for the agents you sell, and how to build an assurance program that keeps pace with a product that changes every quarter. Why AI Assurance Has Become a Sales Requirement for Agent Vendors Security reviews now ask AI-specific questions Enterprise security questionnaires were built for software that behaves the same way every time. Agents do not, and reviewers know it. The questions now arriving in vendor reviews cover model provenance, training on customer data, prompt injection defenses, tool permissions, human approval for high-impact actions and how model updates are communicated. A traditional security program can answer some of these, but rarely with third-party evidence. The gap between what the buyer asks and what the vendor can prove is where deals slow down. Self-attestation stopped being enough For most of the last decade, “we follow best practices” backed by a policy document was an acceptable answer on emerging topics. That changed as AI assurance matured into a recognized discipline with its own standards, registries and certifying bodies. The UK government has published a roadmap to professionalize third-party AI assurance, the Cloud Security Alliance extended its STAR registry to AI services, and AIUC-1 introduced agent-specific certification backed by technical testing. When a credible independent option exists, a self-declared answer reads as a choice not to use it. Procurement teams increasingly treat it that way. The cost of choosing the wrong evidence Assurance is expensive in time more than money. A certification that takes months to achieve and does not answer the buyer’s actual question is worse than no certification, because it consumes the window in which the deal was winnable. Vendors often discover this after the fact: the ISO 42001 certificate satisfies the governance reviewer, and then the security architect asks for evidence that the agent resists prompt injection. Choosing the right instrument starts with understanding what each one proves. That is the subject of the next section. The AI Assurance Instruments Available Today Five types of evidence dominate AI assurance conversations with enterprise buyers. They differ in what they examine, who issues them and how long they remain current. None of them covers everything, which is why mature vendors hold more than one. SOC 2 SOC 2 is an attestation report issued by a licensed CPA firm against the AICPA Trust Services Criteria: security, and optionally availability, processing integrity, confidentiality and privacy. It remains the baseline that almost every enterprise buyer expects from a SaaS vendor, including vendors whose product is an agent. Its strength is that it proves the infrastructure and organization around the agent are controlled. Its limit is that the criteria were not written for AI, so a clean SOC 2 report says nothing about whether the agent can be manipulated into misusing a tool. Elevate’s guide to SOC 1 and SOC 2 consulting explains how readiness for the report works. ISO/IEC 42001 certification ISO/IEC 42001, published in December 2023, is the international standard for AI management systems. Certification confirms that an organization has the policies, roles, risk and impact assessments, lifecycle controls and review cycles needed to govern AI responsibly. It is organization-wide and framework-level, which makes it the strongest answer to “how do you govern AI?” and a natural fit for buyers in regulated industries. It does not require technical testing of a specific agent’s behavior. Elevate’s ISO 42001 executive summary covers the clause structure in more detail. CSA STAR for AI The Cloud Security Alliance launched STAR for AI in October 2025, built on its AI Controls Matrix, a framework of more than 240 control objectives mapped to ISO/IEC 42001, NIST AI 600-1 and the EU AI Act. Level 1 is a self-assessment: the organization publishes its AI-CAIQ questionnaire to the public STAR Registry. Level 2 adds third-party validation by pairing an ISO/IEC 42001 certification with a validated AI-CAIQ. For vendors whose buyers already use the STAR Registry for cloud due diligence, it is a low-friction way to make AI disclosures visible and comparable. AIUC-1 certification AIUC-1 is a certification standard written specifically for AI agents, maintained by the Artificial Intelligence Underwriting Company and developed with a consortium of enterprise security leaders. It combines an accredited audit of controls across six domains (data and privacy, security, safety, reliability, accountability and society) with large-scale adversarial testing of the agent itself, typically 1,000 to 5,000 scenarios. The resulting audit report runs to more than 50 pages and includes red teaming results. Certificates are valid for one year, with technical retests every quarter, and the standard is now part of the CSA STAR Registry. It is the only instrument in this list that certifies the behavior of a specific agent rather than the organization around it. Independent penetration test and red team reports A report from an independent penetration test or AI red teaming engagement is not a certification, but it is often the fastest evidence to produce and the one security architects read most closely. It shows that someone outside the development team attacked the agent, what they found and what was fixed. Its weakness is that it is a snapshot with no standard format, so its credibility depends on the tester’s reputation and the depth of the scope. Elevate’s penetration testing services produce this kind of evidence as a standalone deliverable or as preparation for certification. Instrument What it proves Scope Tests agent behavior How it stays current SOC 2
AI Red Teaming for AI Agents: How to Test What Your Agent Does, Not Only What It Says

AI red teaming for AI agents means attacking an agent the way an adversary would: through its prompts, the content it retrieves, the tools it can call and the permissions it holds. The goal is to learn what the agent actually does under pressure. The stakes changed when models started taking actions. In 2025, a single crafted email was enough to make an enterprise copilot exfiltrate internal files without anyone clicking anything, a flaw rated 9.3 on the CVSS scale. This guide covers what agent red teaming must include, how it differs from a penetration test or a model evaluation, how certification testing is reshaping it, and how to build a program that produces evidence your buyers and auditors will accept. Why AI Red Teaming Has Moved From Research Labs to the Security Review Agents turned output risk into action risk A language model produces text that a person usually reads before anything happens. An agent calls an API, updates a record, sends a message or runs code, often with no human reviewing each step. That shift moves the risk from what the system says to what the system does, and it changes what a test has to prove. A jailbreak that makes a chatbot say something embarrassing is a reputational problem. The same manipulation applied to an agent with write access to a CRM, a payments API or a code repository is a security incident. Buyers now ask for behavioral evidence Enterprise security reviews used to accept policies, architecture diagrams and a SOC 2 report as proof that a product was safe to deploy. For agents, that evidence describes intended boundaries without showing whether the agent respects them when someone pushes past them. AIUC-1 is a certification standard written specifically for AI agents. It closes that gap by pairing an audit of controls with large-scale technical testing of the agent itself, and it produces an audit report that includes red teaming results buyers can request. The standard now sits in the Cloud Security Alliance STAR Registry, and vendors including UiPath, Cursor, Harvey, ElevenLabs and Fin have certified. Whether or not you pursue certification, those announcements have taught procurement teams what a credible answer to “how do you know your agent is safe?” looks like. The cost of a finding someone else discovers A failure your own testing finds is a ticket. The same failure found by a customer’s security team during a review is a stalled deal, and found by an attacker in production it is a disclosure event. Agent vulnerabilities are also unusually cheap to exploit. The attack is written in natural language and often delivered through content the agent reads on its own, such as an email, a support ticket, a web page or a shared document. Elevate’s analysis of agentic applications under attack shows how memory, tools and external content multiply the entry points. The economics favor finding these failures first, on your schedule. How AI Red Teaming Differs From Penetration Testing and Model Evaluation Three disciplines get grouped under “AI security testing,” and they answer different questions. Confusing them is the most common reason a test report fails to satisfy the reviewer who asked for it. Each one is useful, but only one of them tests the behavior of an agent operating inside your environment. Dimension Application penetration test Model red teaming AI agent red teaming Target Infrastructure, APIs, authentication The model’s responses The agent’s decisions and actions across tools, data and channels Core question Can an attacker break the system around the AI? Can the model be made to say something harmful? Can the agent be made to do something it should refuse? Typical findings Misconfigurations, injection flaws, broken access control Jailbreaks, toxic output, sensitive data in responses Goal hijack, tool misuse, privilege abuse, unsafe tool calls Evidence produced Findings with CVSS severity Pass rates on prompt sets Scenario results tied to controls, severity thresholds and retest history The three columns complement each other; none replaces the others. A penetration test can confirm that the API gateway enforces authentication while the agent behind it executes an instruction hidden in a customer email. Model red teaming can show that the underlying model refuses harmful requests in isolation. Yet the same model, wrapped in an agent with broad tool permissions, can be walked into a harmful action one reasonable-looking step at a time. Agent red teaming builds on both, and it answers the question an enterprise buyer actually cares about. Elevate’s overview of the OWASP LLM Top 10 covers the model-level risks this layer inherits. The attack surface is everything the agent can reach An agent’s exposure equals the sum of the credentials, tools, data sources and other agents it can touch. That is why scoping matters more in agent testing than in most security work. An agent with read-only access to a knowledge base and an agent that can issue refunds are different risk classes, even when they run on the same model. A useful scope document lists every tool the agent can call, the permissions attached to each, the content sources it ingests without human review, and the channels through which users reach it. If that list does not exist, the red team will spend its first week building it, and that is worth knowing before you budget the engagement. Non-deterministic behavior changes what a pass means Traditional tests are largely repeatable: an injection flaw either exists or it does not. An agent can refuse an attack nine times and comply on the tenth, depending on phrasing, context length or the state of its memory. Agent red teaming therefore runs large scenario sets, repeats them, and measures failure rates by severity instead of declaring a single pass. Certification testing reflects this: AIUC-1 evaluations typically run 1,000 to 5,000 test scenarios, and Harvey reported passing more than 3,000 unique tests with zero critical failures. Define in advance which failure severities are acceptable at what rate, or the results will be impossible to interpret. What
AI Security Risks in 2026: Top Threats to Watch

AI security risks are the threats that arise specifically from building, deploying, and using artificial intelligence, and they have grown sharply as AI spreads through the enterprise. Traditional security controls were not designed for systems that learn, generate, and increasingly act on their own. This guide covers the top AI security threats organizations face in 2026 and the practical steps to reduce them. What Makes AI Security Risks Different Conventional cybersecurity protects networks, endpoints, and data. AI adds a new attack surface on top of that. The model itself becomes a target, the data used to train it can be corrupted, and a system that generates or acts on outputs can be manipulated through nothing more than its inputs. That is why AI introduces threats that existing controls do not fully address. The Top AI Security Risks The threats below are the ones security and governance leaders should track most closely. Prompt Injection Attackers craft inputs that override an AI system’s instructions, causing it to ignore its guardrails, reveal information, or take unintended actions. It is among the most discussed threats because it requires no special access, only clever text. Data Poisoning Adversaries corrupt the data a model learns from, skewing its behavior in ways that are hard to detect after the fact. Poisoned training data can introduce hidden biases or backdoors that surface only under specific conditions. Sensitive Data Leakage Employees paste confidential information into public AI tools, and that data leaves the organization’s control. This is one of the most common exposures and is closely tied to shadow AI, the use of unapproved tools. Model Theft and Extraction Proprietary models represent significant investment, and attackers attempt to steal them outright or reconstruct them by probing their outputs. The result is lost intellectual property and a copy of the system outside any controls. Shadow AI and Ungoverned Tools Unapproved AI tools running outside IT’s visibility are a risk in their own right, because they cannot be secured, monitored, or governed. Ungoverned adoption multiplies every other threat on this list. Agentic AI and Autonomous Action AI agents that take actions across systems hold real permissions, so a manipulated or malfunctioning agent can cause damage at machine speed. The more autonomy a system has, the higher the stakes when it goes wrong. Third-Party and Supply-Chain Risk Most organizations consume AI through vendors and components they do not control, inheriting whatever weaknesses those suppliers carry. Managing this exposure is a core part of AI vendor governance. How to Reduce Your Exposure Reducing these threats is less about a single tool and more about disciplined practice: Elevate Consult helps organizations build the governance that keeps these threats under control. The ISO 42001 AI Governance Readiness Bundle is a structured starting point. AI Security and Governance Frameworks Recognized frameworks turn scattered defenses into a program. The NIST AI Risk Management Framework names security and resilience as core characteristics of trustworthy AI, and the ISO 42001 standard builds controls for AI risk into a management system. Choosing among them is covered in the guide on AI governance frameworks. How Elevate Consult Helps Organizations Manage AI Risk Elevate Consult helps organizations identify, prioritize, and reduce AI security risks within a governance program aligned to ISO 42001 and the NIST AI Risk Management Framework. The work spans AI inventory, policy, access controls, vendor risk, and the testing that keeps these threats from reaching production. Teams ready to get ahead of these threats can start a conversation with the Elevate team. Key Takeaways Frequently Asked Questions What are AI security risks? AI security risks are threats that arise specifically from building, deploying, and using artificial intelligence, such as prompt injection, data poisoning, sensitive data leakage, model theft, and the misuse of autonomous agents. They extend beyond traditional cybersecurity because the model and its training data become targets in their own right. What is prompt injection? Prompt injection is an attack in which crafted inputs override an AI system’s instructions, causing it to ignore its guardrails, reveal information, or take unintended actions. It is one of the most discussed AI threats because it requires no special access, only manipulated text. How do you reduce AI security risks? Organizations reduce AI security risks by inventorying their AI systems and data flows, governing AI use with policy and approved tools, applying least-privilege access, testing systems adversarially for issues like prompt injection, protecting training data and models, and managing third-party AI risk within a structured program. Is shadow AI a security risk? Yes. Shadow AI, the use of unapproved AI tools outside IT’s visibility, is a security risk because those tools cannot be secured, monitored, or governed. It also multiplies other risks, most commonly the leakage of sensitive data into public AI services. How are AI security risks different from traditional cybersecurity risks? Traditional cybersecurity protects networks, endpoints, and data, while AI security risks add a new attack surface: the model can be stolen or manipulated, training data can be poisoned, and systems that generate or act on outputs can be subverted through their inputs. AI requires controls that conventional security tools do not fully provide.
What a Quality AI Risk Assessment Service Should Deliver in 2026

A quality ai risk assessment service addresses a critical gap: while 78% of companies now use generative AI, only 24% of generative AI projects are secured. In fact, 96% of leaders believe that adopting generative AI makes a security breach more likely. We need detailed ai risk assessment frameworks that deliver more than simple compliance checklists. Effective services provide end-to-end evaluation using proven ai risk assessment templates and identify risks with mitigation strategies across your AI lifecycle. They integrate naturally with your existing ai risk management framework and turn vulnerability into controlled innovation. What Makes an AI Risk Assessment Service Comprehensive in 2026 The ai risk assessment landscape in 2026 demands services built on three foundational pillars: complete lifecycle visibility, inclusive stakeholder participation, and practical integration with governance structures already in place. End-to-End AI Lifecycle Coverage Services must track risks across all six stages defined by OECD and NIST: Plan and Design, Collect and Process Data, Build and Use Model, Verify and Validate, Deploy, and Operate and Monitor. Downstream stages receive notably greater attention than early-stage data practices. Complete services address this gap by balancing coverage across the whole lifecycle. The NIST AI Risk Management Framework, released in January 2023, provides the foundation for this approach. NIST released the Generative Artificial Intelligence Profile in July 2024, which helps organizations identify unique risks posed by generative AI and proposes actions that line up with their specific goals. An ai risk assessment framework requires continuous monitoring and up-to-the-minute threat intelligence to adapt to evolving threats. Organizations protect data integrity, security and availability throughout the whole AI lifecycle, from development through training and deployment. Monitoring systems must detect model drift, which represents an expected operational risk as AI performance degrades with changing conditions. Services deliver value by setting up reliable monitoring systems that track metrics in real time and catch performance degradation before it affects operations. Multi-Stakeholder Risk Identification AI risk assessment involves diverse teams including data scientists, domain experts, ethicists and legal professionals. Risk identification techniques such as scenario planning, threat modeling and impact assessments uncover potential risks that single-discipline approaches miss. Pre-mortem simulation breaks through natural optimism bias by assuming failure has already happened. A two-hour pre-mortem session with diverse stakeholders generates 30-50 high-quality risks that conventional techniques never surface. These extracted risks carry richness that traditional approaches lack and emerge from contextual narratives that capture not just what might fail, but why and how that failure might unfold. Incident pattern mining operates on a simple principle: learn from others’ failures before repeating them. This approach transforms abstract risk frameworks into concrete scenarios by studying AI failures on the ground. The MIT AI Risk Initiative identified 831 mitigations from 13 relevant documents published between 2023-2025. These mitigations fall into four categories: Governance & Oversight Controls, Technical & Security Controls, Operational Process Controls, and Transparency & Accountability Controls. Testing and auditing represented the most commonly mentioned subcategory, with 127 mitigations identified. Integration with Existing GRC Programs Organizations should expand their current enterprise risk, compliance and audit programs rather than creating parallel structures. AI governance demands constant watchfulness as systems learn and evolve, which means risks evolve too. Bias and fairness aren’t just ethical concerns but central risks requiring management in real time. Organizations can use their risk assessment processes for cyber threats, regulatory compliance and operational failures as starting points. AI-specific risks then get added to the mix, including model drift, algorithmic bias and adversarial attacks. AI-powered GRC tools provide up-to-the-minute monitoring by integrating with enterprise systems like ERP, HR and cybersecurity tools already in place. Organizations move from reactive to proactive approaches that meet regulatory demands while optimizing operations. Data governance extends to confirm training data remains free of bias, privacy stays managed during collection, and data provenance gets tracked so organizations can trace the lineage of every piece of data their AI touches. Essential AI Risk Assessment Template Components An ai risk assessment template requires five core evaluation areas that address both traditional IT concerns and AI-specific vulnerabilities that emerge throughout deployment. Data Privacy and Security Risk Evaluation GAI systems raise privacy risks because training requires large volumes of data, which in some cases has personal data. Most model developers do not disclose specific data sources on which models were trained. This limits user awareness of whether personally identifiable information was included. Models may leak, generate, or correctly infer sensitive information about individuals. LLMs reveal sensitive information from the public domain during adversarial attacks. Data memorization poses exacerbated privacy risks even for data present in only a small number of training samples. GAI models may correctly infer PII or sensitive data not in their training data by stitching together information from disparate sources. Wrong or inappropriate inferences can contribute to downstream harmful effects, such as adverse decisions that lead to representational or allocative harms. Security vulnerabilities expand beyond traditional threats. Prompt injection modifies input to make a GAI system behave in unintended ways. Direct attacks craft malicious prompts and indirect attacks exploit LLM-integrated applications by injecting prompts into data likely to be retrieved. Data poisoning allows adversaries to compromise training datasets and manipulate outputs or operation. 91% of organizations recognize that they need to do more to reassure customers their data is being used only for intended and legitimate purposes in AI. Model Bias and Explainability Assessment Bias exists in many forms and can become ingrained in automated systems. AI systems increase the speed and scale at which harmful biases show up. Text-to-image models underrepresent women, racial minorities, and people with disabilities when prompted to generate images of CEOs, doctors, lawyers, and judges. Image generator models produce biased or stereotyped output for various demographic groups. They have difficulty producing non-stereotyped content even when prompts request features inconsistent with stereotypes. Harmful bias stems from training data and causes representational harms or perpetuates bias based on race, gender, disability, or other protected classes. GAI systems may perform differently for subgroups or languages. LLMs perform less well for non-English languages or certain
AI Risk Assessment vs. Avoidable Failures: The Real Cost Comparison for Business Leaders

AI risk assessment becomes critical when you consider that compliance failures cost businesses 15-25 times more than original governance investments. More than 80 percent of AI projects fail, and 95% of organizations faced negative outcomes from their AI initiatives. These artificial intelligence fails translate to major financial damage, especially when 77% of companies lost money directly over two years. You need to know why AI projects fail and put complete risk assessment frameworks in place. Generative AI risk assessment can prevent catastrophic losses. We’ll get into the real cost comparison between proactive assessment and avoidable AI failure in industries of all types. The True Cost of AI Failures: Beyond Regulatory Fines Direct Financial Penalties Under Current Regulations Regulatory penalties represent the most quantifiable expense when artificial intelligence fails to meet compliance standards. The EU AI Act establishes a three-tier penalty structure that escalates based on violation severity. Non-compliance with prohibited AI practices triggers administrative fines reaching €35 million or 7% of global annual turnover, whichever proves higher. Organizations that violate outlined obligations face penalties up to €15 million or 3% of total annual turnover. Supplying incorrect or misleading information to regulatory bodies results in fines of €7.5 million or 1% of annual turnover. These penalties extend beyond EU borders. The legislation affects any organization that deploys AI systems within the 27 member states and creates global compliance pressure. Historical enforcement patterns suggest regulators will not hesitate to impose maximum penalties. The EU collected nearly five billion euros in GDPR violation fines since 2018 and established a precedent for aggressive prosecution of regulatory breaches. Litigation and Legal Defense Expenses Litigation costs compound financial damage from AI failure beyond regulatory fines. Companies faced multiple lawsuits in 2024 for “AI washing,” where they overstated AI capabilities and misled investors. The proposed EU liability directive reduces the burden of proof for harmed parties and makes lawsuits against organizations and their AI systems much easier. Legal exposure extends into unexpected territories. A federal court ruled that AI-generated documents used to assess legal exposure lack attorney-client privilege protection. Independent use of public-facing generative AI tools, particularly outside counsel supervision, may forfeit traditional legal protections and leave materials vulnerable to disclosure during litigation. Organizations face direct legal fees and costs for crisis response that pull executives away from strategic initiatives. System Remediation and Rebuild Costs Remediation expenses quickly exceed original development budgets when AI projects deliver poor results. Projects abandoned before production carry an average sunk cost of $4.2 million. Projects that reach completion but fail to deliver value cost substantially more at $6.8 million on average, yet deliver only $1.9 million in value. Cost-unjustified projects represent the most damaging category at $8.4 million average cost and deliver $3.1 million in value with a payback period of 7.8 years against a typical two-year threshold. AI models that perform poorly require retraining, data cleansing and revalidation. Teams must allocate resources to fix what automation was supposed to streamline. This creates additional expenses in time and labor. Some failures cause irreversible damage. That loss becomes irrecoverable if an AI agent deletes a production server with live customer data when no backup exists. Remediation may not just depend on reversing actions. Without proper controls, nothing remains to reverse. Revenue Loss from Halted Operations Operational disruptions from AI failures create cascading financial damage. Data breaches in 2024 cost companies an average of $4.88 million, with lost business and post-breach response costs contributing substantially. Customers lose faith when AI systems misinterpret data, deny services unfairly or make decisions without transparency. Trust takes longer to rebuild and costs more than the original AI project. Shadow AI proliferation makes tracking and managing risks difficult and forces organizations to invest in expensive compliance updates. Organizations without in-house expertise either pull data scientists into security work or pay premiums for external consultants and divert focus from value-generating activities. The organizational AI confidence deficit that follows failed projects makes leadership risk-averse toward AI investment for 12 to 24 months post-failure and delays future initiatives even when those would have succeeded. Indirect Business Impacts of AI Project Failure Market Valuation Drops and Investor Confidence Single AI failures can trigger catastrophic market value destruction within hours. Google’s Bard chatbot shared one incorrect fact during a demo. The company lost over $100 billion in market value overnight. Zillow faced a $304 million write-down when automated valuations collapsed the housing business built on its Zestimate model. These incidents show how investor confidence evaporates when AI projects fail to deliver promised capabilities. The AI valuation landscape mirrors the dot-com bubble pattern. US equity market capitalization sits at twice GDP, much higher than at the dot-com peak. AI stocks drive the S&P 500’s cyclically adjusted price-to-earnings ratio close to dot-com levels. AI stocks carry a median forward 12-month price-to-earnings ratio of 31x compared to 19x for the broader index. An equity crash like the early 2000s would wipe out around $33 trillion of value and exceed US GDP. Financial stability consequences multiply through several channels. The economy faced a 6.3% peak unemployment rate following the dot-com crash. Unemployment took 47 months to return to previous levels and the S&P 500 needed seven years to recover. Firms that welcome AI show higher revenue per employee, lower marginal costs and stronger earnings outlooks. Those lagging behind face rising relative costs and outdated workflows. This widening AI valuation gap forces companies to either welcome AI or risk obsolescence. Operational Inefficiencies from Failed Deployments Pilot programs consume resources while delivering negligible returns. About 5% of AI pilot programs achieve revenue acceleration, while the vast majority stall and deliver little to no measurable effect on P&L. Organizations report that 46% of projects get scrapped between proof of concept and broad adoption. The percentage of companies abandoning most AI initiatives before reaching production surged from 17% to 42% year over year. Only 16% of AI initiatives achieve scale at the enterprise level. Companies do multiple proofs of concept that amount to impractical science experiments. These inspire awe at
How to Evaluate an AI Risk Management Framework When Choosing a Compliance Partner

Organizations are adopting artificial intelligence faster than ever, and an effective ai risk management framework has become essential. 95% of industry professionals expect GenAI to become central to daily workflows within five years. The global banking sector could see genAI add between $200 billion and $340 billion in annual value each year. Regulators are establishing clearer expectations through frameworks like the NIST AI Risk Management Framework, and selecting the right compliance partner is critical. We’ll explore how to review potential partners based on their ai risk assessment capabilities, ai third party risk management expertise and ai risk governance methodologies. This ensures your organization alleviates AI-related risks. Understanding Your Organization’s AI Risk Assessment Requirements Before you evaluate any compliance partner, you need a full picture of where your organization stands. Map your AI landscape, understand which regulations apply to you, and determine how mature your governance processes are. Current AI Use Cases and Risk Exposure Inventory every AI system and workflow you use. This includes in-house models, third-party APIs, embedded AI in SaaS tools, and any automation that scores, ranks, or recommends. Your AI risk assessment must cover the full spectrum from internal drafting tools to high-stakes decision systems. Risk classification drives everything that follows. Minimal-risk applications like internal search tools with no customer effect require basic oversight. Limited-risk systems such as chatbots that route to human agents need transparency notices and user disclosures. High-risk AI systems trigger extensive compliance requirements under frameworks like the EU AI Act. These include systems used in hiring, credit scoring, healthcare access, education decisions, and identity verification. The gap between AI deployment and board-level oversight remains large. Only 14% of boards discuss AI on a regular basis, while just 13% of S&P 500 companies have directors with AI expertise. 45% of firms have yet to bring AI onto the board’s agenda at all. This disconnect creates governance blind spots that compliance partners must help address. Regulatory Compliance Obligations Your compliance obligations depend on where you operate, what data you process, and how AI systems affect people. The NIST AI Risk Management Framework provides voluntary guidance for incorporating trustworthiness into AI design, development, use, and evaluation. Organizations must think over four core functions: governing, mapping, measuring, and managing. The EU AI Act classifies systems by risk level and imposes specific requirements for high-risk applications. High-risk systems require technical documentation, human oversight, post-market monitoring, and bias testing. Fines for non-compliance can reach up to 7% of global revenue. Even limited-risk tools still require user notices and transparency disclosures. Data privacy regulations add another layer. GDPR Article 35 mandates data protection effect assessments for certain AI processing activities. U.S. privacy laws in Colorado, Virginia, and Connecticut impose similar requirements. About 40% of organizations report experiencing an AI-related privacy incident, often with sensitive data exposure through prompts or integrations. Industry-specific regulations further complicate things. Financial services face scrutiny around credit scoring and algorithmic trading. Healthcare AI must comply with HIPAA alongside medical device regulations. Employment applications trigger equal employment laws under the EEOC. Internal Governance Maturity Level Governance maturity determines how prepared you are to work with a compliance partner. While 80% of large organizations claim to have AI governance initiatives, fewer than half can demonstrate measurable maturity. Organizations at the reactive stage receive AI updates sporadically, often after issues surface. Projects run in silos with minimal feedback loops to leadership. Proactive organizations implement structures for ongoing reporting and mandate periodic performance reviews. They use up-to-the-minute dashboards for risk detection. Mature governance embeds controls directly into AI pipelines. These include automated data validation, bias detection checkpoints, and approval gates requiring lineage verification before deployment. Continuous monitoring for drift with alerts happens in real time. One industrial manufacturer integrated model deployment into its master data management workflow and allowed only certified datasets to train production AI. This single change reduced audit time by 30%. Organizations embedding responsible AI governance see up to 40% higher ROI from AI investments due to reduced rework and audit costs. Understanding your maturity level helps you identify which capabilities to seek in a compliance partner and what gaps need immediate attention. Key Framework Standards to Look For in a Compliance Partner Compliance partners demonstrate their expertise through how they arrange themselves with framework standards that have been around for a while. The frameworks they support reveal their technical depth and implementation experience. They also show how well they know how to guide you through overlapping requirements. NIST AI Risk Management Framework Alignment The NIST AI Risk Management Framework provides structured, risk-based guidance for building and deploying trustworthy AI. The framework was released on January 26, 2023. It was developed through a consensus-driven, open, transparent and collaborative process that included public comments, multiple workshops and stakeholder input. Your compliance partner should demonstrate fluency with all four core functions that structure the AI RMF approach. The Govern function establishes leadership and organizational structures to oversee AI systems. Partners need expertise to help you create governance frameworks that integrate AI risk management into broader enterprise risk strategies. The Map function focuses on how you identify, analyze and evaluate AI-related risks within operational contexts. Look for partners who can place your AI systems in context across technical, social and ethical dimensions. Measure covers risk assessment through both quantitative and qualitative approaches. Good partners use mixed methodologies to understand likelihood and what it all means for AI risks. The Manage function covers ongoing evaluation of AI systems for emerging risks and regulatory compliance. Partners should help you implement continuous monitoring processes that detect drift and novel threats. The AI RMF has a companion Playbook, Roadmap, Crosswalk documents and various Perspectives that partners should reference. NIST released the Generative Artificial Intelligence Profile on July 26, 2024. This profile helps organizations identify unique risks posed by generative AI. Partners unfamiliar with this profile may lack current expertise in GenAI risk management. Crosswalk capabilities matter by a lot. Partners should show you how AI RMF subcategories address ISO 42001 organizational controls, EU