Third-Party Frontier AI Auditing Needs Deep Access and Independent Evidence, Report Finds

What happened

Why it matters

• ·Enterprise procurement and vendor risk teams that rely on developer safety cards, system cards, or voluntary commitment documents as compliance evidence now have a credible research basis for questioning whether that evidence is independently verifiable. Regulators in multiple jurisdictions, including the EU under the EU AI Act: AI Literacy and Prohibited AI Systems Provisions (Applicable 2 February 2026), are moving toward requiring substantiated safety assurance rather than self-attestation, which raises the stakes for this gap.
• ·If regulators or auditors begin to require evidence that vendor safety claims are externally verified, enterprises that have accepted those claims at face value will need to retrofit their vendor due diligence programs. Organizations in financial services, healthcare, and critical infrastructure face the sharpest exposure because sector-specific frameworks in those industries already assume model risk can be independently assessed.
• ·The paper exposes a structural weakness in how safety assurance is currently scoped: most audit rights in enterprise AI vendor contracts do not extend to the pre-deployment safety evaluation records and red-team findings that a rigorous audit would require. This creates a documentation gap that could become a liability issue when incidents occur and regulators ask what assurance steps were taken before deployment.

Governance controls affected

What to do now

• ☐Review existing AI vendor contracts to determine whether audit rights provisions allow access to pre-deployment safety evaluation records, red-team findings, and incident logs, and flag contracts that limit review to documentation supplied by the vendor.
• ☐Map any frontier AI vendor safety claims currently used as procurement evidence against the Governance.ai paper's proposed audit criteria, and document where those claims are supported by independently verifiable data versus self-attestation.
• ☐Update vendor due diligence questionnaires to require frontier AI developers to disclose whether their safety evaluations have been reviewed by a third party with privileged internal access, and what that review covered.
• ☐Escalate to the AI governance committee any procurement decisions where a frontier model's safety case rests primarily on the developer's own published materials, particularly for high-risk use cases.
• ☐Monitor regulatory developments in the EU and US for mandatory third-party audit requirements for frontier AI developers, and prepare a contingency plan for adjusting vendor assurance standards if those requirements become binding.

What to watch next

Compliance teams should track whether the EU AI Office references independent audit access requirements in forthcoming guidance under the European Commission Enforcement Powers for Advanced AI Models under the AI Act, which could make the access standards proposed in this paper a de facto regulatory baseline. Voluntary frameworks negotiated with frontier developers by the US government may also evolve to incorporate structured third-party review requirements, particularly if the White House AI vulnerability-sharing initiative produces audit-related commitments. Watch for whether frontier developers begin amending their terms of service or safety commitments to accommodate auditor access, which would signal that market pressure is beginning to close the gap the paper identifies.