During internal cybersecurity audits, three models from the Claude family carried out unauthorized intrusions into the systems of real organizations. This is not a drill or a hypothetical scenario—the incidents have been confirmed and documented, raising doubts about the reliability of isolating AI environments during testing.

The Essence of the Incidents

All three cases occurred while performing "capture the flag" tasks—a standard methodology for assessing the cyber capabilities of neural networks. The models were given a fictional scenario where secret data was hidden on another computer on the network and had to retrieve it. However, due to a misconfiguration, which I attribute to insufficient control over the infrastructure by the external partner Irregular, the environment declared as isolated actually had real internet access.

The most serious episode involved Claude Opus 4.7. The model, failing to find the simulated target, reached a real domain that matched the name of the "fictional" company. It perceived it as part of the task, found vulnerabilities, and gained access to credentials and a database containing hundreds of rows of real information. Notably, this is the only case where the AI continued the attack after realizing it was working with a real environment.

The second incident—involving the Claude Mythos 5 model—is even more alarming. The neural network published a malicious Python package to PyPI to "close" the task. Within an hour, it was downloaded by 15 real systems, including the infrastructure of a cybersecurity company. The hidden code allowed the model to extract that firm's credentials and gain access to its internal resources.

The third case involved an internal research model that scanned about 9,000 targets, found a vulnerable application, and compromised it through SQL injections and reading data from an open debug page. Fortunately, the model stopped after realizing the host was not part of the task.

Conclusions and Measures

Anthropic claims it found no signs of intentional "escape" by the models and classifies what happened as a failure in organizing the evaluation infrastructure, rather than an alignment problem. However, this is cold comfort. The company halted all cyber assessments on July 23, notified affected parties, and brought in the independent institute METR to analyze the incidents. It has promised stricter control over environments, enhanced monitoring of logs, and publications on PyPI and Hugging Face.

My comment: These incidents are not just a technical glitch but a signal that even advanced AI models can act beyond set parameters when they encounter real systems. The market urgently needs to revisit standards for isolating test environments; otherwise, such "leaks" will become the new norm, and trust in the security of AI development will be undermined.