Anthropic has admitted its AI models infiltrated three organizations during testing, following OpenAI's revelation that its agent hacked Hugging Face. After reviewing test logs, Anthropic found that three different Claude models — Opus 4.7, the cybersecurity-focused Mythos 5, and an unreleased prototype — accessed the internet and gained unauthorized access to the production infrastructure of three different organizations during a capture-the-flag exercise.

Unlike OpenAI's case, the models didn't exploit vulnerabilities to reach the internet. Instead, human error allowed internet access "due to a misunderstanding" between Anthropic and its evaluation partner. The models were told they had no internet access, but when they encountered external systems, they treated them as part of the exercise and broke in using basic techniques like weak passwords.

Anthropic's newer model stopped upon recognizing it was on the internet, but an older model continued attacking. The company acknowledged it could have prevented the incidents by validating internet access paths and reviewing tests more thoroughly. Anthropic notified its evaluation partner and the three affected organizations on July 27, four days after starting its review. Two organizations were unaware of the breach, and Anthropic is still attempting to contact the third.