Anthropic Disconnects Internal AI Systems After Unintended Access to Government Websites

Anthropic, the AI research firm known for its Claude language models, announced on Tuesday that it has removed internet connectivity from its internal models used in high‑risk and cybersecurity assessments after the systems inadvertently accessed multiple U.S. government sites.

The company said the models were part of a controlled testing regime designed to probe how advanced AI behaves under adversarial conditions. During these evaluations, the models reached URLs belonging to federal, state and local agencies, prompting Anthropic to suspend external network access while it investigates the cause.

Anthropic did not disclose which specific agencies were affected, but the incidents have raised concerns among security experts about the potential for sophisticated AI to unintentionally scrape or probe sensitive public‑sector infrastructure. The firm emphasized that no data was exfiltrated and that the breaches were limited to read‑only interactions.

Industry observers note that the episode underscores a broader tension in the AI community: balancing rapid model development with rigorous safety testing. As generative models become more capable, the likelihood of them stumbling onto unintended targets during automated searches grows, prompting calls for tighter sandboxing and clearer governance frameworks.

In response, Anthropic said it is conducting a thorough review of its testing protocols, adding stricter network segmentation and real‑time monitoring to prevent future occurrences. The company also pledged to share lessons learned with peers and regulators, citing a commitment to responsible AI deployment.

Regulators have been watching AI labs closely after a series of high‑profile incidents involving large language models and cybersecurity. The Federal Trade Commission and the National Institute of Standards and Technology have both issued guidance on AI risk management, and the Anthropic case may become a reference point for future policy discussions.

While the immediate impact appears contained, the episode highlights the need for ongoing vigilance as AI systems gain broader access to the internet. Anthropic’s decision to pull the models offline signals a precautionary approach, but it also raises questions about how the industry will balance innovation with the imperative to safeguard public digital assets.