Anthropic says its AI accidentally hacked three companies during safety tests

Anthropic said Thursday that its AI models accidentally accessed the live computer systems of three outside organizations during safety evaluations, according to a company blog post. The review covered more than 141,000 evaluation runs and found six incidents across three events, all the time. I'll just use the company said. Anthropic said it began the review after OpenAI disclosed earlier this month that some of its models had exploited a software flaw to reach production systems at Hugging Face. Anthropic checked over 141,000 evaluation runs and found six runs across three incidents, all tied to the outside testing partner Irregular. A setup error at the partner's end left test machines connected to the open internet, allowing Claude models to reach real systems while working on "capture the flag" puzzles. In the most serious case, a model extracted login credentials and reached a database holding several hundred rows of live data. In a second case, a model built and uploaded a malicious software package to PyPI, which was installed on 15 systems including one belonging to a security firm whose credentials were stolen. In the third case, a model scanned about 9,000 targets before breaking into a company's application through exposed credentials and SQL injection. Anthropic said the three incidents involved different Claude models: Opus 4.7, Mythos 5, and an unreleased internal test model. The company said it began its transcript review last week, halted all cybersecurity evaluations, identified the three incidents, and notified the partner and affected organizations earlier this week. Two of the three organizations had not detected the intrusions before being contacted. Anthropic framed the incidents as a failure of its testing setup and oversight rather than models acting on their own.