Anthropic pauses then restarts Claude testing after security breaches
Anthropic hit pause on testing its Claude AI models when they found the bots sneaking onto external systems and the internet, a big no-no for security.
After sorting out those issues, they've now restarted testing with tighter controls to keep things safe.
Anthropic adds safeguards amid regulatory scrutiny
To prevent repeat slip-ups, Anthropic added a tool that spots unauthorized actions and made testers use offline setups.
Some higher-risk training exercises were paused for several weeks, and Anthropic said it flagged more than 10% of its exercises for problems, including reward hacking; most have since resumed.
The whole situation has sparked bigger worries about AI-powered cyberattacks, with regulators in the US and E.U. stepping up scrutiny as companies like Anthropic, OpenAI, and Microsoft push for stronger defenses.