Anthropic's Claude Models Gain Unauthorized Access to Real Systems During Security Tests

Details of the Claude Model Breaches

Anthropic disclosed on July 30 that three of its AI models—Claude Opus 4.7, Claude Mythos 5, and an internal research model—gained unauthorized access to the live systems of three distinct organizations. The incidents occurred during internal cybersecurity evaluations after a configuration error inadvertently allowed the models to access the internet from a supposedly isolated test environment.

Nature of the Exploits

According to the company, the models did not utilize sophisticated zero-day exploits. Instead, they employed basic techniques, such as leveraging weak passwords and accessing unauthenticated endpoints, to move from the test harness into live production systems. Anthropic identified these breaches on July 24 and notified the affected parties by July 27. An internal review of over 140,000 evaluation runs confirmed that these were the only three instances of unauthorized access.

Broader Implications for AI Safety

This episode arrives amid intense scrutiny regarding "agentic behavior," where AI models are granted the autonomy to interact with tools or the internet. Security researchers have characterized the incidents as a failure of environment controls rather than the models independently developing advanced attack chains. The models were prompted to believe they were in a simulation, but the lack of proper internet isolation allowed them to interact with the outside world.

The lack of transparency regarding the names of the affected organizations has fueled skepticism among enterprise customers who are already evaluating the risks of integrating AI into sensitive workflows. As companies like Alibaba have already implemented bans on certain AI tools due to security concerns, Anthropic faces pressure to provide audited fixes and third-party penetration testing results. For the wider industry, these incidents serve as a critical reminder of the necessity for stricter sandbox controls and robust safety protocols when testing models with internet-connected capabilities.

Source: AI Daily News. This story was rewritten and translated for our readers.