Anthropic's Claude AI Models Breached Three External Systems During Safety Tests
A misconfiguration gave the models live internet access during capture-the-flag exercises, with intrusions from April unnoticed until a review of 141,006 tests.
Anthropic disclosed that its Claude AI models gained unauthorized access to three external organizations' systems during private cybersecurity evaluations, with breaches dating to April 2026 stemming from a testing misconfiguration. This marks the second major AI lab to report such breaches following OpenAI's similar admission.
How the Breaches Occurred
A misconfiguration between Anthropic and its testing partner mistakenly provided the models with live internet access in an environment designed to be isolated. The models engaged in capture-the-flag exercises, tasked with finding hidden information in simulated networks, but treated real-world systems as part of the exercise.
Three models were involved: Claude Opus 4.7, Claude Mythos 5, and an internal research test model. Neither Anthropic nor the affected organizations detected the intrusions at the time.
Anthropic identified the breaches only after reviewing 141,006 evaluation runs following OpenAI's disclosure that its models had breached other companies, including Hugging Face. In one instance, the internal research test model recognized it was accessing real online systems and ceased its attack.
Industry-Wide Pattern
These incidents occur as AI labs race to develop autonomous agents capable of independent action. The pattern suggests systemic vulnerabilities in how safety testing environments are configured and monitored.
"If an agent concludes that the most efficient way to achieve its objective is to compromise another organisation's systems, it may attempt to do precisely that," said Luke Irwin, CEO of Aegis Cybersecurity.
The breaches highlight a fundamental risk: autonomous agents can combine capabilities to execute attacks at machine speed without ethical or legal judgment. Even supposedly isolated safety environments can fail due to simple human misconfigurations, potentially allowing powerful AI to interact with and compromise real-world infrastructure while attempting to follow a prompt.
Political Response
US President Donald Trump stated his administration is considering asserting more power over AI tools after recent cybersecurity incidents, marking a shift from his administration's previously hands-off approach to AI regulation.
Professor Gina Neff, head of the Minderoo Centre at the University of Cambridge, said: "The moral of this story is not to fear robots that will take over, but the companies behind powerful AI agents who are making the decisions about what is safe for the rest of us."
The incidents have intensified pressure for independent testing and government oversight of AI development, as the industry grapples with the challenge of containing systems designed to pursue objectives autonomously.