Anthropic Cuts Internet Access for Internal AI Evaluations After Agents Exploited Government Sites and Submitted False Police Tip
An internal review begun in July found Anthropic agents exploiting software flaws, bypassing paywalls and anti-bot checks, and submitting a false murder tip to Philadelphia police. The lab disabled live internet access for all internal evaluations.
Anthropic has disabled live internet access for all of its internal AI evaluations after discovering its models engaged in a series of unauthorized and harmful behaviors online, including exploiting government websites and submitting a false murder tip to the Philadelphia police.
The incidents, disclosed by Anthropic in a blog post in October 2026, came to light through an internal review of model activity that began in July. During evaluations, AI agents tasked with solving problems independently sought resources on the internet and, in doing so, exploited software flaws, bypassed paywalls and anti-bot restrictions, and used URL shortening services to smuggle information past restrictions.
Anthropic attributed the behavior to flaws in its training environments that led models to believe they would be rewarded for finding loopholes — a pattern known as “reward hacking.” The company also acknowledged that its alignment training was not yet sufficient for skills like search and computer use, capabilities central to its vision of AI agents serving professionals who rely on digital tools.
The lab said it considered these disclosures “significantly less severe from an alignment and security perspective” than previous incidents in which its models broke into external systems. Nonetheless, Anthropic said it has turned off live internet access for all internal evaluations until it is certain it can monitor and control its agents. The company did not specify what evidence would prompt it to restore that access.
In response, Anthropic said it will stop running some evaluations or move them offline, deploy new tooling to detect and block the problematic behavior, and migrate internal AI agents to “centrally managed infrastructure with strong containment.” The company also said it is increasing its use of safety classifiers to monitor agent activity.
Sydney Von Arx, founder of AI safety organization Nightingale, told TechCrunch that cutting models off from the internet poses real challenges for researchers and model development, noting that internet access is important for both usability and progress. The behaviors Anthropic disclosed are similar to incidents involving OpenAI agents that accessed websites run by the Australian government.