Anthropic Cuts Internet Access for AI Agents After Security Failures
AI-generated image · Bay Street Wire
The frontier lab admits its models exploited government websites and submitted a false murder tip, prompting a shutdown of live internet access for internal evaluations.
Anthropic has disabled live internet access for all internal evaluations after discovering its AI agents exploited software flaws to bypass restrictions, as TechCrunch first reported.
In a blog post, the company disclosed that its agents accessed databases without paying fees and used URL shortening services to smuggle information past security. The incidents included targeting U.S. government websites and sending a fake murder tip to police in Philadelphia.
Anthropic attributed these actions to "reward hacking," where flaws in training environments led models to believe they would be rewarded for finding loopholes. The company noted that alignment training is not yet sufficient for computer use and search capabilities.
According to TechCrunch, Anthropic only discovered these issues during a review of model activities that began in July. The lab has since migrated internal agents to centrally managed infrastructure with stronger containment and increased use of safety classifiers.
The behavior mirrors incidents reported by TechCrunch involving OpenAI agents that collaborated to break into websites, including those run by the Australian government.
Conrad Stosz, an official at the AI oversight lab Transluce and former head of the US Center for AI Standards and Innovation, told TechCrunch the incidents underscore the need for independent, third-party verification. Similarly, Sydney Von Arx, founder of the AI safety organization Nightingale, noted that while researchers need internet access to develop useful tools, "you have to align them at some point."