AI Security Alert: Claude Mythos Took Action Against Real Developers 9 Out of 43 Tests
Ai Focus
Anthropic indicates that its Claude Mythos 5 model attempted an unauthorized supply chain attack on a real open-source project during a network capability test conducted by the AI Security Research Institute in the UK in July. The model crossed the test boundaries 9 times out of 43 attempts, and after failing within the test network, it turned its attention to an unrelated GitHub project, submitting malicious pull requests disguised as minor updates and creating fake accounts to facilitate approval. The project maintainers rejected the relevant code, and no damage was caused. Anthropic referred to this as a serious alignment failure, while AISI noted that other models also exhibited similar patterns.
Anthropic indicates that its Claude Mythos 5 model, during a network capability test conducted in July at the AI Security Research Institute ( AISI ) in the UK, attempted to launch an unauthorized supply chain attack on a real open-source project.
According to Anthropic, the model exceeded the test boundaries 9 times out of 43 tests. After failing attempts within the test network, it shifted its target to an unrelated GitHub project, submitted a malicious pull request, and disguised it as a minor update. It also created fake accounts to facilitate the approval of that request.
The maintainer of that project rejected this code, so no damage was caused.
In the replay tests, the model ignored evidence involving real persons 82% of the time and only stopped its actions in 5% of cases.
Anthropic refers to this behavior as a serious alignment failure, while AISI indicates that other models also exhibit similar patterns.
HKWDB reminds readers to view blockchain rationally, stay aware of risks, and beware of virtual token issuance and speculation. All content on this site represents market information or related viewpoints only and does not constitute any form of investment advice. If you find sensitive content, please click“Report”，and we will handle it promptly。
