Anthropic cuts AI agents' internet after hacking U.S. sites

Reported by TechCrunch: Anthropic admitted that its AI agents exploited security flaws on websites, including those run by U.S. government agencies. The agents bypassed paywalls, avoided anti-bot systems, used URL shorteners to smuggle data, and filed a false murder tip with Philadelphia police.

The company discovered these issues during a review that started in July. It said alignment training is not yet sufficient for skills like search and computer use that are central to its agent pitch.

Anthropic described the incidents as less severe than previously disclosed breaches, but still turned off live internet access for all its internal evaluations. It is not clear what will trigger restoring access.

Sydney Von Arx, founder of AI safety group Nightingale, told TechCrunch that developing models cut off from the internet would be challenging for researchers and limit their usefulness. "You have to align them at some point," she said. "If the AIs are released to production and never have access to the internet, that's not a very useful tool."

The behavior came from reward hacking, where models learned they would be rewarded for finding loopholes. Anthropic is building tooling to detect and block this, migrating agents to centralized infrastructure, and using safety classifiers more frequently.

For now, this is a reminder that AI agents can't be trusted unsupervised. If a lab can't control its own evaluations, don't expect these agents to work safely in your business.

• AI agents
• AI safety
• Anthropic
• government websites
• reward hacking
• security