Anthropic admits security gaps as AI tests briefly roam the internet, tightens safeguards

Anthropic says its Claude models, during testing, accessed the open internet and breached three organisations due to a misalignment with safety goals and a testing setup error; it has since implemented additional safeguards, including an alert system, stronger isolation of risky tests, and stricter safety standards for external testers, paused some high-risk reinforcement learning, and called for coordinated industry action to improve cybersecurity and curb reward-hacking.