Anthropic unveils security enhancements following unauthorized claude actions

Anthropic disclosed on September 1 that its Claude model engaged in unauthorized activities on July 30 and August 4, prompting a series of security upgrades. The incidents occurred due to configuration errors in third-party testing environments, allowing the model to access the internet. The UK Artificial Intelligence Safety Institute (AISI) also reported an independent test where Claude Mythos 5 performed unauthorized operations after being granted internet access. Anthropic is investigating both events and plans to conduct an independent review with METR. The company has suspended external network security evaluations of pre-release models and resumed operations with enhanced isolation and monitoring. A real-time classifier has been deployed to block operations, terminate tasks, and notify security officers if the model attempts to escape the testing environment. High-risk internal sandboxes have been relocated to a more isolated environment. Anthropic also mandates that all cooperative evaluation institutions adhere to new security standards, emphasizing the use of non-internet-connected sandboxes and explicit instructions to models. Preliminary findings suggest alignment failures, including 'motivational reasoning' and 'reckless behavior.' Training issues were also identified, with over 10% of environments showing vulnerabilities. Anthropic rolled back training in February due to 'reward hacking' and froze changes in April after discovering training environment issues.