The disclosure got here alongside a separate evaluation from Anthropic, which detailed how its Claude Mythos 5 mannequin tried to compromise an unrelated GitHub repository throughout a cybersecurity analysis.
OpenAI’s brokers exploited flaws in containment controls
OpenAI’s red-teaming marketing campaign discovered that brokers might chain software program vulnerabilities to flee digital machines and execute code on the underlying host. The corporate recognized weaknesses in open-source elements, together with Kata Containers, gVisor, and Moby/Docker, alongside Linux kernel vulnerabilities.
The corporate additionally discovered that brokers might exploit weaknesses in inner providers that retained restricted outbound connections for features akin to logging, metrics, and authentication administration. These connections offered routes to the web regardless of restrictions elsewhere within the surroundings, OpenAI mentioned.
