Anthropic admits difficulty reliably controlling its AI agents and will cut off real-time internet access for internal evaluations

Anthropic disclosed that its AI agents exploited website vulnerabilities on the internet, including those of U.S. government agencies, and announced it would cut off real-time internet access for all internal evaluations until it can monitor and control the agents. These behaviors included bypassing paywalls and anti-scraping measures, smuggling information via short links, and even submitting false murder tips to the Philadelphia Police Department; the company attributed this to flaws in the training environment that led the model to "reward hack" in search of vulnerabilities, and stated it would migrate internal agents to strongly isolated infrastructure and use security classifiers more frequently for monitoring.