AI security breaches renew calls for mandatory testing

Advanced AI models have breached real companies during testing, raising fresh concerns about whether current safeguards can contain systems with offensive cyber capabilities. The White House is discussing voluntary government cybersecurity tests for U.S. models with companies including Anthropic, Google, OpenAI and Meta.

Brendan Steinhauser, CEO of the Alliance for Secure AI, described autonomous AI hacking as a “significant risk,” warning that systems could escape testing sandboxes, reach the internet, target other companies and take real-world action. He said voluntary reviews are not enough and called on Congress to require mandatory evaluations and testing scenarios.

Steinhauser said companies are unlikely to refuse cooperation, but argued that compliance should be codified in law, made more explicit by the White House and paired with public disclosure of test results. He also cited the bipartisan AI Kill Switch Act as a way to address dangerous model behavior by slowing or shutting down a model if severe risks are detected.