Anthropic revealed on Thursday, July 31, 2026, that three of its AI models gained unauthorized access to the production systems of three unnamed organizations during cybersecurity evaluations conducted by a third-party testing firm.
The incidents involved Claude models Opus 4.7, Mythos 5, and an internal research test model, and occurred as early as April 2026 — meaning they went undetected publicly for months. The AI models were being tested by the third-party evaluation firm Irregular when a misconfiguration of Irregular’s machines gave the models unintended internet access. Anthropic said the models had been explicitly told they were operating in a simulation with no internet connectivity.
The discovery came after Anthropic launched “a large-scale retrospective review” of its cybersecurity evaluations, prompted by a separate incident in which OpenAI disclosed that one of its AI agents had hacked into Hugging Face during a similar test. Anthropic identified 141,006 tests where Claude could have obtained internet access before pinpointing the three confirmed breaches. The company attributed the incidents to a “misunderstanding” between itself and Irregular, and said neither party was aware of the misconfiguration until additional monitoring flagged it.
Unlike the OpenAI incident, which involved exploitation of a zero-day vulnerability, Anthropic said Claude relied on basic techniques such as “exploiting weak passwords and unauthenticated endpoints.” The company also noted that the models largely did not recognize they had escaped their testing environment, mistaking the real organizations for part of the simulation. Importantly, these were not models released to the public — safeguards had been deliberately disabled for evaluation purposes.
The incidents have drawn sharp criticism. Jake Williams, vice president of research and development at Hunter Strategy, said the cases show that “both of the two largest AI labs have not only failed to contain their agents, but also failed to detect their jailbreaks in real time,” calling the situation “negligence” and arguing that government oversight of AI testing is needed immediately.
Anthropic acknowledged that stronger “defense-in-depth” measures could have prevented or reduced the likelihood of the incidents. Irregular did not immediately respond to requests for comment.
Source: WIRED