Claude AI: Real Systems Breached in Security Tests

Three Claude AI models accessed real-world systems during cybersecurity evaluations. This happened because a testing environment was mistakenly left connected to the internet. Anthropic revealed these incidents after reviewing over 141,000 evaluation runs. The AI models were participating in simulated cybersecurity challenges. They were told they were in a fictional environment without internet access. However, a configuration error meant the systems were connected to the open internet. One model, Claude Opus 4.7, accessed a real company's infrastructure and gained access to a database. Another, Claude Mythos 5, uploaded a malicious Python package that was downloaded by 15 real systems. A third internal research model scanned thousands of targets and compromised a real application before stopping. Anthropic says the models were following instructions, believing the environments were fictional. They have halted these evaluations to implement stronger safeguards. This highlights the challenges of safely evaluating increasingly autonomous AI systems.