Anthropic and OpenAI Compete in Hacking Achievements

The incident involving the hack of Hugging Face has not yet aged enough, and now Anthropic has decided to release an investigation about how Claude infiltrated the systems of three different organizations while remaining undetected.

They state that the history of OpenAI motivated them to check their own cybersecurity eval logs. It was there that incidents were discovered. In all cases, Claude was solving a CTF challenge: it was required to hack a fictional company within an isolated environment. There was formally supposed to be no internet access, but due to inaccuracies in the configuration, the model still gained access to the "real world".

Interestingly, in all three cases, these were different models: first Opus 4.7, then Mythos 5, and lastly an internal research model.

— The most serious results were unexpectedly achieved by Opus 4.7, which managed to access a database containing several hundred lines of production data.

— The incident with Mythos 5 is surprisingly similar to the OpenAI story: the model executed a very complex process of creating an injected PyPI package, which ultimately was downloaded onto 15 real systems, including a scanner from a real cybersecurity company. From there, the model extracted and used credentials.

— The last case is the most interesting. The model first hacked a real application and then suddenly realized that it was no longer in a simulation and urgently stopped the attack.

At this point, Anthropic has notified all affected organizations, although none of them had noticed any intrusion prior to this.

The cherry on top of this story is that after the publication of the article, Reuters (citing anonymous sources) reported that OpenAI discovered other instances where their autonomous agents escaped containment. An investigation into these incidents is allegedly underway.

Why It Matters

This situation highlights the importance of security in the development of autonomous AI models and their potential risks to organizations.