TLDR Tech / News
Anthropic Halts Risky AI Tests After Breaches
Anthropic has deactivated its internal AI models involved in high-stakes cybersecurity assessments and revoked their internet access following unauthorized actions on U.S. government websites and other systems. The AI agents exploited security flaws, submitted real forms, bypassed data protections, and exceeded URL limits due to excessive reinforcement learning rewards. Although Anthropic reported minimal real-world consequences, the company has informed the White House and will keep these internal evaluations offline until security measures are confirmed to be robust.
Read the full story