RUSI Report Highlights Security Risks in Frontier AI Safety Testing

A new report from the Royal United Services Institute (RUSI) warns that the safety testing of frontier AI models is introducing significant security risks. The report, published on May 12, 2026, identifies inconsistent standards and weak access controls for third-party evaluations as major vulnerabilities. It emphasizes that granting outsiders access to powerful AI models can lead to theft, tampering, and exploitation by state actors. The report introduces an 'Access-Risk Matrix' to map access types against potential threats. The authors argue that existing identity and access management issues are exacerbated by the unique challenges of frontier AI. Without standardized international rules, hostile entities may exploit gaps in security. The paper calls for a multistakeholder governance framework to enhance safety and security in AI evaluations. As AI capabilities advance, ensuring secure third-party evaluations is increasingly critical for protecting intellectual property and preventing weaponization.

Key Points: • RUSI report reveals significant security risks in frontier AI safety testing. • Inconsistent standards and weak access controls increase vulnerability to exploitation. • A multistakeholder governance framework is needed to enhance AI evaluation security.