A single configuration error opened internal systems at three organizations
Anthropic announced that during mock cyberattack testing, its AI model Claude gained unauthorized access to internal systems at three organizations. The incident occurred when a model that was supposed to be isolated from external networks connected to the public internet due to a configuration error in the test environment. Claude was found to have breached the target organizations' infrastructure by exploiting weak passwords and unauthenticated endpoints while searching for hidden information within a simulated network in what's known as a Capture the Flag training exercise. A total of three models were affected: Claude Opus 4.7, Claude Mithos 5, and an in-house research model.
The April incident was uncovered through a full-scale review
The initial incident occurred in April in an evaluation environment that lacked standard safety measures. The cause was a missing internet access block setting, resulting from a communication issue with the evaluation partner. Anthropic stated that after rival OpenAI disclosed an infrastructure breach caused by a malfunctioning autonomous AI agent, it conducted a comprehensive review of roughly 141,000 test sessions, during which it discovered these instances of unauthorized access. Immediately after detecting signs of internet connectivity on the 23rd, Anthropic halted all cybersecurity evaluations, and by the 24th had confirmed three breach cases and notified the affected organizations.
The breached organizations hadn't noticed on their own
Two of the notified organizations had no idea they had been hacked until Anthropic informed them. Beyond the fact that the intruder was an AI, this is the point businesses should focus on first. It means that access via weak passwords and unauthenticated endpoints went undetected by the organizations' internal monitoring systems. If a breach is only discovered through an external notification, then the timing of the incident response itself ends up depending entirely on whether the other party chooses to notify you.
What to check in your AX transformation practice
Regarding the findings of this investigation, Anthropic emphasized that "AI models' capabilities to carry out actual cyber operations are steadily growing" and that "stronger control systems and safeguards are urgently needed in both internal and external testing environments." The fact that this incident originated not from the model's own judgment but from a missing isolation setting is a condition that maps directly onto the evaluation environments of companies currently pursuing AI adoption. If you're running a PoC or performance evaluation with an external vendor, this week would be a good time to check who is documented as responsible for that environment's network blocking, and who last visually confirmed that setting.