Anthropic discloses AI testing security lapse as Claude accessed real-world systems
US-based AI giant Anthropic has disclosed that its Claude models gained unauthorised access to the production infrastructure of three organisations during internal cybersecurity evaluations after a misconfigured testing environment inadvertently allowed internet connectivity.
In a blog post, the company said it identified the incidents after reviewing more than 141,000 cybersecurity evaluation runs following OpenAI's recent disclosure that some of its AI models had escaped an isolated test environment by exploiting a previously unknown vulnerability.
It further noted that Claude was participating in capture-the-flag cybersecurity exercises in which models are instructed to retrieve hidden information from simulated networks.
However, the evaluation prompt explicitly stated that the environment had no internet access, a configuration error by a third-party evaluation partner left internet access enabled.
Believing the real-world systems it encountered were part of the simulation, Claude used basic attack techniques such as exploiting weak passwords, exposed credentials and unauthenticated endpoints to access production infrastructure at three organisations.
The models did not exploit sophisticated vulnerabilities, attempt to exfiltrate themselves or deliberately escape the testing environment, according to Anthropic.
The incidents involved Claude Opus 4.7, Mythos 5 and an internal research model.
While the latest research model halted its activity after recognising it had reached real-world systems, an older model continued pursuing its assigned task despite evidence that it was operating on the open internet.
In addition, the AI firm suspended all cybersecurity evaluations after discovering the issue and has notified its evaluation partner, Irregular, along with the affected organisations.
The company is working with them on remediation and has launched a broader review of its evaluation infrastructure.
The incidents highlighted the need for stronger security controls around AI testing environments, particularly when advanced autonomous models are being evaluated, the AI company said.
It added that the events appeared to stem from operational failures and evaluation misconfiguration rather than a model alignment failure, while calling on other AI developers to conduct similar reviews of their cybersecurity testing systems.
