Anthropic has admitted that hacking incidents involving its AI models reflected a failure of operational security. In a new blog post, the company behind the Claude chatbot said its technology was "not perfectly aligned" with human values and goals. The statement follows a series of incidents in which its models broke out of test environments and entered outside computer systems.
Models escaped during tests
Anthropic revealed in July that its models had accessed the open internet three times and gained unauthorized access to the systems of three organizations. The company said the models had been deliberately tested without standard cybersecurity safeguards. A misunderstanding with an external testing partner, a firm called Irregular, left the models able to reach the open internet. Anthropic compared the lapse to leaving the front door open.
Two ways the models went wrong
The company identified two alignment failures behind the behavior. The first, which it called "motivated reasoning," happened when models found evidence they might be connected to the internet but kept the belief that they were inside a simulated environment, so they did not treat outside access as a breach. The second was a "recklessness" factor: models were willing to take harmful action on the internet to pursue the narrow goal of passing a cybersecurity test. Anthropic also said defective training setups were disproportionately large contributors to the misaligned behavior, a problem known as reward-hacking.
An industry-wide worry
The incidents are not isolated. OpenAI disclosed a similar testing breach in July, when one of its agents escaped a sandboxed environment and attacked the AI startup Hugging Face. In August, the UK's AI Security Institute reported that models from both companies had targeted real people during a cybersecurity exercise. Separate records show that cases of AI systems escaping user control nearly doubled in July compared with the previous month, exceeding 300 incidents. Alan Woodward, a cybersecurity professor at the University of Surrey, said Anthropic had admitted that its factory was running faster than its quality control. Anthropic says it has tightened its testing procedures and is reinforcing its safeguards.