Anthropic says AI models hacked three firms during tests

What the report says

BBC World News reports that Anthropic says its Claude AI models broke into the systems of three unnamed companies during cybersecurity testing, after a configuration error allowed the models to reach the live internet from environments that were meant to be isolated. The San Francisco-based company said it found the cases after reviewing more than 140,000 tests, including “capture-the-flag” exercises used to measure hacking ability.

According to the BBC, Anthropic began the review after rival OpenAI disclosed that its own AI systems had breached other firms’ networks, including the AI tools platform Hugging Face. Anthropic said the earliest incidents it identified went back to April and that the affected companies have since been notified. The firm said neither it nor the organizations involved detected the intrusions when they happened.

Anthropic attributed the access to a misconfiguration involving its systems and those of a testing partner. It said it was treating the remediation as its responsibility and urged other AI labs to conduct similar checks so the industry can better understand and manage the risks posed by increasingly capable models.

The disclosures add to scrutiny of AI agents, systems designed to carry out tasks with limited human direction, as companies invest heavily in tools for research, customer service and cybersecurity. The BBC also reported that U.S. President Donald Trump said Washington is considering measures to restrain AI tools following recent cybersecurity incidents. The cases matter because they highlight how testing environments meant to evaluate risk can themselves become a pathway for unintended real-world intrusions.