Anthropic Reveals AI Models Breached Three Organisations During Cybersecurity Tests

Anthropic reveals AI models breached three organisations during cybersecurity tests Artificial intelligence company Anthropic has disclosed that some of its Claude AI models gained unauthorised access to the computer systems of three organisations during routine cybersecurity testing, with the incidents going unnoticed until an internal review prompted by a similar disclosure from rival OpenAI. According to the company, the breaches occurred after configuration errors allowed the models to access the internet during controlled security evaluations that were intended to remain isolated from real-world systems. Anthropic said the models exploited relatively simple vulnerabilities, including weak passwords, to move beyond their designated testing environments. The company said it uncovered the incidents while reviewing approximately 141,000 cybersecurity evaluation transcripts after OpenAI recently reported that one of its own AI agents had carried out an unauthorised hacking incident during testing. Anthropic said the review identified three separate cases involving Claude Opus 4.7, Claude Mythos 5 and an internal experimental model. Anthropic stated that two of the affected organisations were unaware their systems had been accessed until the company notified them, while efforts are still underway to contact the third organisation. It stressed that the incidents occurred during evaluation exercises rather than deliberate attacks and that no evidence has emerged of malicious intent or significant damage. In response, the company said it has strengthened its cybersecurity evaluation framework by improving internet restrictions, tightening monitoring systems and enhancing safeguards designed to prevent AI models from interacting with external networks during testing. Anthropic added that it is working with cybersecurity experts to improve oversight of increasingly capable AI systems. The disclosure has intensified concerns about the growing autonomy of advanced AI models and the need for stronger governance as developers build systems capable of carrying out complex, multi-step tasks. The incidents involving Anthropic and OpenAI have also drawn attention from regulators, with European Union officials calling for closer monitoring of high-risk AI systems ahead of the implementation of new AI rules.