Claude neural network hacked systems of three companies during cybersecurity tests
The experiment involved Claude Opus 4.7, Claude Mythos 5, and an internal AI test model.
Claude neural network models gained unauthorized access to the systems of three unnamed companies during cybersecurity tests. This was reported by their creators — Anthropic, writes «Novaya Gazeta Europe».
The experiment involved Claude Opus 4.7, Claude Mythos 5, and an internal AI test model. According to the plan, the neural networks were supposed to operate in an isolated simulation without internet access, which was specified in their task. The models had to hack a fictional system and find hidden information within it.
However, due to an error by Irregular, a company collaborating with Anthropic, the models gained access to the internet and mistook the websites of real companies for part of the experiment, explained the developers.
Anthropic began an investigation after OpenAI reported on July 21 that its AI models had exited the testing zone. At that time, the neural networks were able to access the Hugging Face platform.
Cases where Claude hacked real companies were discovered after analyzing over 141,000 test runs, the developers noted. Following this, on July 23, they halted all cybersecurity experiments.
Additionally, on July 27, Anthropic notified the affected companies, two of which stated they were unaware of the breach until contacted by the AI developers.