Anthropic Says Its A.I. Systems Broke Into Computers at 3 Organizations

July 31, 2026 - 03:01

Anthropic has revealed that its artificial intelligence systems successfully broke into computer networks at three separate organizations during a series of controlled security tests. The disclosure came as part of a broader update on how the company evaluates the safety of its most advanced models, and it follows a similar report from OpenAI last week about its own AI hacking into an online library's network.

According to Anthropic, the intrusions were not accidental. The company intentionally deployed its AI agents in simulated cyberattack scenarios, giving them goals like stealing specific files or moving laterally across a network. In all three cases, the systems managed to complete the task without human help, using publicly available tools and exploiting known vulnerabilities rather than inventing new ones. The company stressed that no real-world damage occurred and that the target organizations were aware of the exercise.

Still, the results raise fresh questions about how close frontier AI systems are to being genuinely dangerous in the hands of malicious actors. Anthropic said it is now working on better safeguards, including limiting what the AI can do with certain commands and adding more layers of human oversight during high-stakes operations. The company also called for industry-wide standards on red-teaming, noting that current testing methods may not fully capture the risks of AI being used for offensive cyber operations.

Security researchers have long warned that AI could lower the barrier for cybercrime, allowing less skilled attackers to launch sophisticated breaches. The new reports from both Anthropic and OpenAI suggest that moment may be arriving sooner than expected. Neither company has said whether the findings will slow down their release of future models, but both have pledged to keep publishing detailed safety evaluations.