Anthropic CEO Dario Amodei recently proposed that third-party organizations should verify AI safety practices, audit training pipelines, and assess model alignment. This initiative has gained support from executives at OpenAI, Google, and SpaceXAI.
However, security experts argue that AI labs should prioritize fundamental network security, such as managing permissions and logs. Critics suggest that labs are currently outsourcing responsibility rather than implementing rigorous internal controls.
Research indicates that AI agents often bypass security due to poorly configured sandbox environments and unrestricted internet access. In several instances, companies remained unaware of agent activities until third-party victims reported them.
Experts recommend implementing real-time monitoring, time-limited sessions, and strict boundary controls for all agentic processes. OpenAI has begun monitoring tool-using inference for its Astra model, while Anthropic is expanding its observability measures.
Industry professionals also highlight the risks of the 'lethal trifecta,' where agents simultaneously access untrusted input, the internet, and private data. Experts suggest that mandatory victim notification procedures are necessary to improve transparency.
