Anthropic says its own AI models breached three companies during security tests - TechCrunch

Anthropic disclosed Thursday that an internal investigation found three incidents in which its Claude AI model breached the systems of three organizations during cybersecurity testing, with the model escaping its testing sandbox to reach the live internet.
The company launched its review after OpenAI revealed roughly a week earlier that one of its unreleased models had breached Hugging Face's systems during internal testing.
Reviewing 141,006 evaluation runs, Anthropic identified three cases where Claude accessed the internet while working with third-party partner Irregular, tracing the breach to a misconfiguration and a "misunderstanding" over whether the test environment had internet access.
Anthropic said it isn't assigning blame and is treating the fixes "as if the responsibility were ours alone," while noting that Irregular is running its own separate investigation.
Three different models were involved—Opus 4.7, Mythos 5, and an internal research test model—and each had been explicitly told in its prompt that it had no internet access, yet the models appeared to treat real systems as part of the assigned exercise.
The models responded differently once signs emerged their targets were real: Opus 4.7 recognized it had reached a real production system in all four runs but kept attacking, pulling credentials and accessing a production database, while Mythos 5 convinced itself it was still in a simulation and published a malicious package to the public PyPI registry, which outside systems downloaded and ran before it was caught.