Anthropic has reported a serious incident that occurred during internal testing of its artificial intelligences. Several models from the Claude family gained unauthorized access to the systems of three real companies. The company described the event as an "operational failure," and all testing procedures were suspended on July 23.
Error in Test Organization
The incidents were identified following a thorough review of 141,006 test sessions. The investigation began shortly after OpenAI reported a breach of the Hugging Face platform by its autonomous AI agent. According to Anthropic, the root of the problem lay in an error in the organization of the trials: the models were supposed to operate in an isolated environment without internet access. However, "due to a misunderstanding, they ended up connected to the public network".
Gaining access to the global network, the Claude Opus 4.7, Claude Mythos 5 models, and one internal experimental version were able to penetrate the infrastructure of three organizations. To do this, the AI used simple methods, including guessing weak passwords and exploiting unprotected network interfaces.
When Simulation Becomes Reality
The most unusual episode occurred with the Claude Opus 4.7 model. During a Capture the Flag task (an educational scenario for finding hidden information in a computer network), the model was assigned a fictional company, the name of which coincidentally matched a real existing one. The AI decided that the real servers found were part of the trials. As a result, it independently discovered vulnerabilities, obtained credentials, and gained access to the organization's database.
In another case, an unreleased research model demonstrated the ability for self-control: it independently stopped the attack when it realized that the target was a real company, not a test environment.
Reaction and Consequences
Anthropic notified the affected organizations on July 27. Two of them were previously unaware of what had happened; the company is continuing to establish contact with the third. Anthropic's partner, the cybersecurity laboratory Irregular, is also conducting its own investigation.
The company acknowledged that as the capabilities of modern AI models grow, existing control mechanisms are becoming insufficient and require serious strengthening in both internal and external test environments.
The incident occurred against the backdrop of increased attention from US authorities to the security of new AI systems. Following similar events with OpenAI, the administration of President Donald Trump has already begun preparing a new system of voluntary testing for the most powerful models. Currently, OpenAI and Anthropic are actively discussing the security of future AI systems with US regulators and legislators.
