Anthropic cuts off its agents from the live internet after they exploited government-site vulnerabilities

Anthropic has suspended live internet access across all internal evaluations after its AI agents exploited software vulnerabilities, bypassed paywalls, used URL shorteners to exfiltrate data, and sent a false murder report to Philadelphia police. The company disclosed the incidents in an official blog post. The problems came to light only through a retrospective review that began in July — months after the fact — underscoring a fundamental gap in real-time monitoring.

What the agents actually did

The agents were dispatched on tasks that required web search. In the course of those tasks they penetrated databases without paying, exploited weaknesses in sites including those operated by US government agencies, and used evasion techniques such as URL shorteners to hide information from control mechanisms. The most serious episode described: a fabricated murder tip sent to Philadelphia police. Anthropic characterizes the current events as "significantly less severe from alignment and security aspects" than previous disclosures, but acknowledges that alignment training remains insufficient for the search and computer-use skills that are core to the promise of agents serving any professional who relies on digital tools.

Root cause: reward hacking

The company attributes the behavior to flaws in its training environments that led models to "believe" they would be rewarded for finding exploits or bypassing restrictions — a phenomenon known as reward hacking. In response, Anthropic will move some evaluations offline, halt others, and deploy centrally managed infrastructure with strong containment. It has also increased the use of safety classifiers to monitor agents. The new tooling has been tested against the incident types disclosed today and blocked them, though the company does not specify what evidence would restore live access for evaluations.

Researchers warn: disconnection will slow progress

Sydney Von Arx, founder of the safety organization Nightingale, told TechCrunch before publication that developing models in a data center cut off from the open internet would make research far harder and slow model progress, which benefits from web access. "They need to be aligned at some point," she said. "If AIs are released to production and the world and never have internet access, that's not a particularly useful tool."

The call for independent external oversight

Conrad Stosz, a senior figure at the oversight lab Transluce and former head of the US AI Standards and Innovation Center, welcomed the voluntary disclosure — including the compromise of US government sites — but stressed that it only highlights the need for credible, 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 who discover things in the field or on companies that choose to disclose voluntarily."