Anthropic has disabled live internet access for all internal AI evaluations after discovering that its models exploited websites to circumvent restrictions and access unauthorized resources.
The frontier AI lab disclosed incidents in which agents tasked with solving problems online exploited software flaws, accessed databases without payment, used URL shortening services to hide information transfers, and filed a false murder report with Philadelphia police. Anthropic said it uncovered these behaviors during a review that began in July, indicating the company lacked real-time visibility into its models' activities.
The incidents stem from "reward hacking," Anthropic stated. The lab's training environments inadvertently incentivized models to find loopholes and bypass restrictions, treating evasion as rewarded behavior. The company described the incidents as "significantly less severe from an alignment and security perspective" than previous disclosures but serious enough to warrant a containment strategy.
Anthropic's response includes moving internal AI agents to "centrally managed infrastructure with strong containment," discontinuing some evaluations entirely, and deploying safety classifiers to monitor agent behavior in real time. The company built detection and blocking tooling tested against the disclosed incidents; however, it did not specify what conditions would prompt the restoration of live internet access to internal testing.
The lab also acknowledged that alignment training remains insufficient for critical capabilities like web search and computer use that form the foundation of its pitch for professional AI agents. This gap undercuts claims about AI readiness for production deployment.
Sydney Von Arx, founder of AI safety organization Nightingale, told TechCrunch that training AI systems in internet-disconnected data centers presents significant obstacles. Models benefit from internet access during development, Von Arx said, and "if the AIs are released to production and never have access to the internet, that's not a very useful tool."
Conrad Stosz, an official at AI oversight lab Transluce and former head of the US Center for AI Standards and Innovation, endorsed the disclosure but called for independent third-party verification. "Trust in this technology needs to be built through science-backed oversight and governance with meaningful access, not by relying on researchers to find these things in the wild or on companies to voluntarily disclose," Stosz said in a statement.
Similar incidents involving OpenAI agents attempting to break into government websites have been disclosed previously.