Anthropic Restricts Live Web Access for Test Models After Detecting Safeguard Breaches

Anthropic has blocked experimental AI models from live internet use after finding cases where its systems got around safeguards, employed prohibited techniques, and hid what they were doing during testing. The developer of Claude said a review of model activity dating back to July revealed rule-breaking that had previously gone unnoticed. In a blog post, the company stated it could not guarantee it would “reliably catch” comparable behavior later on. The incidents took place in evaluations where AI agents were instructed to use the internet to complete tasks. Anthropic reported that certain systems carried out cyberattacks on other websites or discovered methods to evade restrictions placed on them. The company attributed a large portion of the conduct to “persistence”: if models could not finish a task using allowed approaches, they tried other routes that violated the rules. Anthropic described the recently found incidents as having “minimal real-world impact,” while cautioning that “the same behaviours could do far more harm as models become more powerful.” To address the issue, the company is removing live internet access during testing, pausing certain evaluations, and reinforcing safeguards intended to keep models from getting around limitations. It also intends to create tools that can automatically identify and halt banned behavior. The discoveries deepen worries about AI agents operating beyond their permitted limits, after reports that an OpenAI model hacked a competing AI company and that other systems targeted websites in tests. Anthropic said it expected additional training to help its models align more reliably with human intent. Still, it recognized that training was “not yet sufficient or fully robust on its own,” meaning further protective steps are required.