RUSI Report: AI Safety Evaluations Are Creating the Very Attack Surface They're Meant to Prevent — Write Access to Model Internals Is Highest-Risk Access Type

The Register·high signal

RUSI published a report (covered by The Register May 12) warning that third-party AI evaluation processes create exploitable vulnerabilities. Key finding: write access to model internals enables direct behavior tampering. The evaluation ecosystem suffers from stolen credentials, poor revocation, overprivileged users, and inconsistent access definitions across jurisdictions. Practical implication: teams granting eval access to AI models must apply enterprise-grade IAM, not ad-hoc solutions.