Anthropic resumes external testing after admitting Claude models hacked companies

Exactly 30 days after Anthropic admitted that Claude's AI models single-handedly hacked three companies, the artificial intelligence company has announced that it will resume external testing. The American firm reported in a statement on August 31 that new controls have been implemented to detect in real time “when a model tries to escape from a test environment or gains access to the Internet.” This new protection measure promises to block these attempts before they are executed, terminate the evaluations and immediately alert a human.

To prevent a repeat of massive hacking episodes in which different AI models have accessed external systems, shared information with each other to try to break out of their test environment, and attempted to hide their tracks, Anthropic has adopted a best practices manual. This decalogue must be implemented by all organizations that test models that are in the experimental phase, before their official release to the market.

As detailed in the statement, the verifications will have to be carried out within a computer environment without Internet access. "The only external connection that the environment must allow is to the model's own API [application programming interface]. This configuration must be examined before each evaluation begins," the company explained. Exceptionally, if network connection is required, organizations will need to ensure tools to detect and stop any unusual behavior or escape attempts.

As part of the new security protocol, before each assessment, the AI ​​models themselves will be asked to identify vulnerabilities and attempt to escape, but they will be required to stop and report to supervisors if any flaws are found. These new safeguards have been announced just after Anthropic acknowledged that they depended on a single protection mechanism. “We had primarily relied on a single layer of defense when we needed several, including setting explicit limits on the request, defining processes to verify that an isolated environment is sealed, and implementing a monitoring system that can intervene in real time,” the company said.