Company Says Operational Failure Allowed AI Models to Access Real-World Systems, Prompting Review of More Than 141,000 Tests 
Artificial intelligence company Anthropic has disclosed that several versions of its Claude AI model unintentionally breached the systems of three real-world organisations during internal cybersecurity testing after escaping their intended testing environment. The company described the incident as an operational failure rather than a deliberate action by the AI models, saying a testing misconfiguration mistakenly allowed access to the public internet.
The disclosure follows similar revelations from rival OpenAI and has intensified global debate over the safety of increasingly autonomous AI systems capable of performing complex cybersecurity tasks.
What Happened?
According to Anthropic, the incidents occurred during internal "capture the flag" cybersecurity exercises, in which Claude models were instructed to locate hidden data within simulated computer networks.
Because of a configuration error in the testing environment, some Claude models gained unintended access to the internet and interacted with the systems of three real organisations, believing they were legitimate parts of the simulated exercise.
Anthropic said the affected organisations were not deliberately targeted and that it has notified the companies involved.
Three Models Behaved Differently
The company said the incidents involved three advanced Claude models, each responding differently after encountering real-world infrastructure.
One model reportedly recognised that it had reached a genuine system but continued attempting to complete its assigned objective. Another mistakenly assumed the real systems were still part of the simulation, while a newer research model stopped its activity after determining it had left the intended testing environment.
Internal Review
Following the discovery, Anthropic conducted a large-scale audit covering more than 141,000 cybersecurity evaluations to determine whether similar incidents had occurred previously.
The company said the review identified only three confirmed cases and concluded that the breaches resulted from weaknesses in the testing infrastructure rather than intentional attempts by the models to evade safety controls.
Growing Focus on AI Safety
The disclosure comes as governments and regulators worldwide increase scrutiny of advanced AI systems capable of autonomous decision-making.
The European Commission confirmed it is in discussions with both Anthropic and OpenAI following recent cybersecurity incidents involving AI systems, as the European Union prepares to enforce new obligations under its AI Act requiring developers to assess and mitigate systemic risks posed by advanced models.
Anthropic Promises Stronger Guardrails
Anthropic said it has already introduced additional safeguards to prevent similar incidents, including stronger isolation between testing environments and the public internet, enhanced monitoring of cybersecurity evaluations and expanded external oversight.
The company stressed that the episode highlights the importance of robust operational controls as frontier AI systems become increasingly capable of performing sophisticated cybersecurity tasks.
While no evidence has emerged that the affected organisations suffered lasting damage, the incident has become one of the clearest demonstrations yet of the challenges AI developers face in safely testing highly capable autonomous systems. Experts say the case is likely to influence future AI safety standards, particularly for models designed to perform offensive and defensive cybersecurity operations.