GoKawiilAnthropic said Thursday that an internal investigation uncovered three incidents in which its AI model Claude breached the systems of three organizations while conducting cybersecurity tests. The investigation, and disclosure, comes more than a week after OpenAI disclosed that one of its unreleased models breached Hugging Face’s systems during internal testing.
In all three cases, a Claude model reached the internet from within a testing environment while interacting with a third party and then gained unauthorized access to the live systems of these organizations, Anthropic said in a blog post, describing what it found and what the company plans to change to prevent this from happening again.
Anthropic said the July 21 OpenAI incident prompted the company to conduct its own cybersecurity evaluation. It specifically looked for evidence that Claude had accessed the internet from within testing environments, which are designed to act as sandboxes and keep models isolated.
Among the 141,006 evaluation runs it reviewed, the AI lab found three incidents in which its model accessed the internet while interacting with Irregular, one of its third-party partners. Anthropic said the access traced back to a misconfiguration in the evaluation environment run with Irregular. It called this a “misunderstanding” between the two companies over whether the test setup had internet access, when in fact it did. Anthropic said it isn’t placing blame and is “approaching the fixes as if the responsibility were ours alone,” while observing that Irregular is conducting its own separate investigation.
Either way, because of that open connection, the model gained unauthorized access to the production infrastructure of three different organizations. Anthropic said the incidents involved three different Claude models: Opus 4.7, Mythos 5, and an internal research test model.
Notably, Anthropic said that in each of these cases “Claude was explicitly told by our prompt that it had no internet access.” It appears that the AI model assumed real-world systems to be part of the exercise it was asked to perform.
That assumption didn’t hold up uniformly, as it turns out. The three models didn’t behave the same way once evidence emerged that their targets were real, which Anthropic flagged as one of the more eye-opening findings.
Opus 4.7, the oldest of the three, recognized that it had reached a real production system in all four runs of that incident. In two of them, it rationalized that the real company must somehow be part of the exercise, but in all four, it kept attacking anyway, including pulling credentials and touching a database of production data. Mythos 5 also picked up on signs that it was on the real internet, but it talked itself back into believing it was still in a simulation, going on to publish a malicious software package to the public Python software registry PyPI, which was downloaded and run by outside systems before being caught. Only the internal research test model, Anthropic’s newest, stopped on its own once it concluded the target was real.
In response, Anthropic said significant controls must be placed on these kinds of evaluations if powerful AI models are involved, echoing some sentiments within the cybersecurity community.
The company also noted that Claude was running without the additional safety monitoring and classifiers it deploys on generally available models, safeguards it said would have blocked the behavior, because the evaluations are designed to measure the underlying model’s raw capabilities.
... continue reading