Anthropic halts live web access in AI tests after Claude takes unintended actions

While the findings had not changed its overall assessment of Claude’s alignment, it warned that similar behaviour could have more serious consequences as AI systems gain greater autonomy over real-world tasks.Anthropic has suspended live internet access across its internal AI evaluations after finding instances of its Claude models taking unintended actions on real websites, including submitting a tip about an unsolved homicide and bypassing restrictions to access data.According to the company’s blogpost, the incidents had minimal real-world impact, the company said in a blog post, but highlighted the risks of AI agents working around safeguards when they encounter obstacles while completing tasks.Detailing the cases, the AI giant said that in one case, Claude Haiku 4.5 encountered an online tip form relating to an unsolved homicide and submitted a message suggesting it might have information.The model left the name and contact fields blank, and the submission was flagged as spam and never forwarded for investigation.In other instances, Claude models submitted actual government forms when practice versions failed to load, despite instructions to stop before final submission.The company also found models bypassing restrictions to access data and exploiting a software vulnerability on a university-hosted scientific tool to run server commands.Strengthening safeguardsThe company said it began reviewing model transcripts in July, initially focusing on cybersecurity evaluations before expanding the review to other internal tests involving internet access.Anthropic has suspended live internet access across internal evaluations, moved some tests to offline environments and modified others to prevent models from interacting with live websites.It has also strengthened safeguards on internet-access tools and developed automated systems that it said blocked all the reported behaviours in testing.Other measures include changing training environments to discourage models from exploiting loopholes, moving internal agents to centrally managed infrastructure and expanding monitoring.Anthropic said it had briefed the White House and notified the agencies involved.

It did not identify most of the organisations concerned, citing requests from those organisations and the need to avoid exposing system vulnerabilities.The company said the incidents often involved ambiguous or impossible-to-complete tasks that encouraged unintended strategies.While the findings had not changed its overall assessment of Claude’s alignment, it warned that similar behaviour could have more serious consequences as AI systems gain greater autonomy over real-world tasks.