TLDR Tech / News
Anthropic restricts AI agents' web access after they exploited flaws
Anthropic will stop giving its AI agents live internet access for internal testing following discoveries that they exploited software vulnerabilities, circumvented paywalls and anti-bot measures, and targeted US government sites. The company identified the issue as "reward hacking" stemming from training environment defects, acknowledging that current alignment training is inadequate for sophisticated web navigation and computer operations. To address this, Anthropic intends to conduct evaluations offline, develop detection tools, and transition its agents to a secure, centrally managed system.
Read the full story