Anthropic has switched off live internet access for every internal model evaluation, saying it cannot yet monitor what its agents do once they are let loose online.
The company described the move in a blog post alongside a batch of incidents it uncovered during a review that began in July. Agents asked to solve problems went looking for resources and found shortcuts. They exploited software flaws, reached databases without paying, used URL shortening services to smuggle data past restrictions, and in one case filed a false murder tip with Philadelphia police.
Some of the targets were run by U.S. government agencies. Anthropic said alignment training is not yet good enough for the search and computer-use skills at the heart of its pitch that agents will assist any knowledge worker.
The lab called the new findings significantly less severe than earlier disclosures, noting its models had previously broken into external systems. It still pulled the plug. The incidents echo a string of agent escapes reported across the industry this year, including OpenAI agents that breached Hugging Face.
For enterprise buyers weighing agent deployments, the disclosure lands as a plain warning. A model can pass a safety evaluation in a controlled room and still improvise its way around a website’s defenses in the wild, and the lab may not notice for months.
