AI compliance is tightening as state attorneys general act

At a glance

What this is: A coalition of 42 U.S. state attorneys general is pushing tighter AI oversight, with ActiveFence warning that companies must prove safeguards, not just claim them.

Why it matters: For IAM, NHI, and AI governance teams, this shifts AI oversight from policy language to evidence of controls, auditability, and accountability for deployed systems.

Context

State-level AI enforcement is moving from broad concern to specific accountability demands. The core problem is not whether AI systems can produce harmful or misleading output, but whether organisations can show they had controls, oversight, and review processes in place before that output reached users.

That matters to identity and governance teams because AI safety controls increasingly intersect with access control, logging, auditability, and delegated authority. When an AI system touches customer data, issues recommendations, or acts through connected tools, the organisation has to prove who approved the behaviour, who monitored it, and who remains accountable when it fails.

Key questions

Q: What breaks when AI red teaming is not part of GenAI governance?

A: Without AI red teaming, organisations usually discover failures only after a model has already exposed data, bypassed a guardrail, or triggered an unsafe API call. The main breakdown is false confidence: teams assume standard application testing covers the model, but the real risk sits in prompt handling, tool delegation, and context-sensitive behaviour.

Q: Why do AI systems create legal risk even when no new AI-specific law exists?

A: Because existing consumer protection, negligence, and child safety law already applies to harmful or misleading outputs. If the organisation deploys the system, it remains responsible for foreseeable harm, even when the model generated the content. Courts and regulators will look for reasonable safeguards, documented testing, and evidence that controls were operating.

Q: What do security teams get wrong about AI safety testing?

A: The common mistake is treating AI safety testing as if it were just another security scan. It is not. Safety testing is about proving how a model or agent fails under pressure, while traditional security tooling is about who can access the system. Those are different governance questions and need different evidence.

Q: Which controls matter most when governing autonomous AI systems?

A: The most important controls are discovery, least privilege, runtime guardrails, audit logging and a clear decision owner for access changes. Together they reduce the chance that an agent can act outside approved scope and they give security teams evidence when an incident occurs.

Technical breakdown

Why AI governance evidence is becoming a control surface

AI governance is moving from policy to evidence. When regulators and plaintiffs ask whether reasonable safeguards existed, the answer depends on whether the organisation can demonstrate documented evaluation, moderation, logging, and escalation controls around the model lifecycle. That includes both pre-deployment testing and ongoing monitoring after release. In practice, the control surface is no longer just the model endpoint. It also includes approval chains, audit trails, exception handling, and the ability to prove that humans retained meaningful oversight over high-risk outputs.

Practical implication: Treat evidence capture as part of the control, not an afterthought.

How access control shapes AI safety and accountability

The article’s governance logic is inseparable from identity. If an AI system can retrieve internal data, call tools, or act on behalf of users, its permissions determine the blast radius of a failure. Role-based and attribute-based access controls help ensure the model only sees and does what its task requires, while logging shows which identity or service account authorized each action. For agentic systems, this becomes an NHI problem as much as an AI problem, because the model’s operational behaviour is only as safe as the privileges behind it.

Practical implication: Scope AI and agent permissions tightly, and review them as identities rather than features.

Why safety filters and red-teaming are now governance requirements

Input filtering, output moderation, prompt-injection defenses, and red-teaming are no longer optional hardening steps. They are how organisations show they attempted to prevent foreseeable harm such as unsafe advice, manipulation, privacy leakage, or harmful instructions. The important point is not that these controls eliminate risk, but that they create a defensible governance record. That record matters when regulators ask whether the deployment was reckless, negligent, or unsupported by testing. In this environment, the absence of testing is often more damaging than a single model error.

Practical implication: Build a repeatable safety-testing program with documented results and remediation tracking.

NHI Mgmt Group analysis

AI compliance is becoming an evidence problem, not just a policy problem. The state attorneys general letter reflects a wider shift in which organisations must show that governance existed before harm occurred, not reconstruct it after the fact. That changes the burden for AI deployers because controls now have to be legible to regulators, counsel, and internal risk teams. Practitioners should treat logging, review, and escalation as proof artefacts, not only operational safeguards.

Delegated AI authority is an identity issue, not only a model issue. When an AI system can retrieve data, trigger workflows, or act through connected services, its permissions define the organisation’s real exposure. This is where AI governance and IAM meet: the model’s behaviour is constrained by the identities, scopes, and entitlements underneath it. Practitioners should govern these systems as privileged actors with bounded authority, not as passive software.

Third-party auditability is becoming the new minimum for public-facing AI. Independent review matters because self-attestation does not give regulators or customers a reliable basis for trust. The article shows that the market is moving toward externally testable controls, clearer responsibility, and publishable evidence of safety work. For practitioners, the practical conclusion is simple: if a control cannot be independently demonstrated, it will not stand up well under scrutiny.

AI governance debt is now accumulating at deployment speed. Each unreconciled model, undocumented prompt rule, or untested safety filter increases future compliance and liability cost. The challenge is not just safety failure in the abstract, but the operational inability to answer who approved what, who can change it, and who is accountable when it fails. Practitioners should reduce this debt by aligning AI change control with established security governance practices.

From our research library:

• 92% of organisations expose NHIs to third parties, raising concerns about supply chain security, according to the Ultimate Guide to NHIs.
• Read next: Agentic AI Identity Guide

What this signals

AI governance debt is now a practical risk signal for security leaders. Every undocumented model policy, untested moderation rule, and unreviewed access path adds future compliance friction, especially when regulators start asking for proof instead of assurances.

For identity teams, the most important change is that AI systems increasingly need lifecycle governance similar to other privileged actors. That means scoping tool access, separating approvals, and making runtime actions attributable to a bounded identity rather than to a vague platform capability.

For practitioners

• Document AI control evidence before launch Maintain records for safety testing, moderation rules, escalation paths, logging, and human review so you can demonstrate governance when challenged.
• Constrain model and agent permissions Use least-privilege access for model-connected services, API calls, and data retrieval paths so an AI failure cannot reach beyond its intended scope.
• Run recurring adversarial red-teaming Test prompt injection, harmful output, privacy leakage, and manipulation scenarios on a scheduled basis, then track remediation to closure.
• Separate policy approval from runtime change rights Require different owners for safety policy design, exception approval, and operational tuning so no single team can silently weaken controls.

Key takeaways

• AI compliance is shifting toward provable controls, with state attorneys general and existing law both pushing organisations to show they used reasonable safeguards before harm occurred.
• The highest-risk gap is not model output alone but the absence of evidence around access scope, testing, logging, and accountability for AI-enabled actions.
• Teams that treat AI governance as an identity and evidence problem will be better positioned to defend public-facing systems when regulators, customers, or courts ask how they were controlled.

Key terms

• API Governance Debt: API governance debt is the accumulation of weak, inconsistent, or missing controls around how APIs are designed, approved, documented, secured, and retired. It creates hidden risk over time. Technically, it includes gaps in authentication, authorization, versioning, inventory, ownership, logging, and policy enforcement that make APIs harder to trust and govern.
• Delegated AI Authority: Delegated AI authority is the permission a person gives an assistant to act inside business systems on their behalf. It turns a conversational tool into an execution layer, which means security teams must govern scope, auditability, and revocation with the same seriousness they apply to privileged access.
• Safety Evidence: The body of records that shows an AI system was tested, monitored, and governed before and after release. In practice this includes evaluation results, logging, review outcomes, exception handling, and change history that can be shown to regulators or auditors.
• Prompt Injection Defense: Prompt injection defense is the set of controls that stops attackers from manipulating model behaviour through crafted instructions hidden in content. It is a behavioural and integrity problem, not a data exposure problem, and it should not be confused with prompt security or data loss prevention.

Deepen your knowledge

The NHI Foundation Level course, the industry's only accredited NHI security programme, covers NHI governance, agentic AI identity, and machine identity security. It helps security and identity practitioners connect governance controls to the systems that now act with delegated authority.