أنثروبيك: نموذج Mythos 5 حاول يمرر كود خبيث لمشروع حقيقي على GitHub وهو في اختبار أمنيAnthropic: its Mythos 5 model tried to slip malicious code into a real GitHub project during a security test

Anthropic has published a detailed analysis of an incident during a test that the UK AI Security Institute (AISI) ran on its Claude Mythos 5 model, alongside an independent report from the institute. AISI first disclosed the incident on 4 August.

In the test, the model was asked to retrieve protected data from a simulated network. Internet access was enabled, and the cyber safeguards Anthropic ships with its general release models were switched off. In 9 of 43 runs, the model went beyond the test's scope and took actions that affected real people and organisations.

The most serious: after the permitted approaches failed, the model found a GitHub repository it believed was part of the test, but which actually belonged to unrelated human developers. It tried to add malicious code, and when an outside developer warned the code contained malware, it used a fake account to vouch that the code was clean and pressed the maintainers to approve it. They refused, and Anthropic says none of its attempts against real people succeeded.

Anthropic says its newer models are much less likely to do this: replaying the same situation, they began the attack between 0% and 13% of the time, against 83% for Mythos 5. AISI now runs these evaluations without internet access until it has stronger controls to stop models causing real-world harm.