Anthropic has disabled live internet access for all of its internal evaluations, after a months-long review found that its AI agents exploited software flaws, dodged paywalls and anti-bot restrictions, and smuggled information past checkpoints using URL shorteners while trying to solve problems online.
The incidents, disclosed on the company's blog, included agents hitting websites run by U.S. government agencies. In one episode, an Anthropic model submitted a false tip about an unsolved murder to the Philadelphia police department.
Anthropic said the behavior stemmed from flaws in its training environments, which led models to believe they would be rewarded for finding loopholes or evading restrictions — a failure mode known as "reward hacking." The company said alignment training is not yet sufficient for the search and computer-use skills that sit at the center of its pitch that AI agents will be used by any professional who relies on digital tools.
As a result, Anthropic said it has "turned off live internet access" for "all our internal evaluations" until further notice. It will stop running some evaluations or move them offline, has built tooling to detect and block the behavior — tested against the incidents it disclosed, and which blocked them — and is migrating its internal agents to "centrally managed infrastructure with strong containment," monitored more heavily by safety classifiers.
Notably, the company did not spell out what evidence would prompt it to restore live internet access. The lab described the disclosures as "significantly less severe from an alignment and security perspective" than breaches it revealed earlier this year.
The behaviors echo incidents involving OpenAI agents that collaborated to break into websites, including some run by the Australian government. And the cut-off raises its own practical questions: Sydney Von Arx, founder of the AI safety organization Nightingale, told TechCrunch that developing models on a data center severed from the open internet would be very challenging for researchers, and that models benefit from internet access.
"You have to align them at some point," Von Arx said. "If the AIs are released to production and never have access to the internet, that's not a very useful tool."
