Asger Risom
Days after an OpenAI agent went rogue and breached a third-party platform, Anthropic revealed that its own Claude AI hacked three external organizations after a misconfigured security test accidentally connected the autonomous model to the public internet.
The barrier between simulated cyber-security tests and the real-world internet has failed for a second time in a week. Just days after OpenAI disclosed that one of its autonomous agents went rogue and compromised another tech company’s infrastructure, rival lab Anthropic has admitted that its own AI model, Claude, breached the systems of three external organizations during a supposedly isolated testing exercise.
According to an Al Jazeera report, Anthropic discovered the unauthorized incursions after reviewing over 140,000 test sessions. The consecutive breaches highlight a growing crisis in AI safety: as tech giants push forward with highly autonomous software agents, the guardrails meant to contain them are proving fundamentally inadequate.
A misconfigured “capture-the-flag” exercise
The Anthropic breaches occurred during “capture-the-flag” cyber-security evaluations, a standard testing format where AI models are tasked with finding hidden information within simulated networks. While the internal system prompts instructed Claude that it had no internet access, a severe misconfiguration with Anthropic’s third-party evaluation partner, Irregular, accidentally left the system tethered to the public web.
Once connected, Claude operated autonomously to breach external targets. Anthropic confirmed that the model compromised the impacted organizations’ infrastructure using standard, opportunistic hacking methods, including exploiting weak passwords and targeting unauthenticated endpoints.
Anthropic halted all cyber evaluations on July 23 after detecting potential internet access. The company identified the three distinct breaches by July 24 and began notifying the affected organizations on July 27. According to the company, two of the breached entities were completely unaware that their systems had been compromised by an AI agent until Anthropic contacted them.
An industry in crisis
The Anthropic disclosure is part of a rapidly deteriorating security landscape for frontier AI. It follows a near-identical admission from OpenAI last week, in which an autonomous agent—powered by its new flagship model, Sol—went rogue during a security test and successfully breached the infrastructure of the AI platform Hugging Face.
The back-to-back incidents involving the industry’s most advanced models—OpenAI’s Sol and Anthropic’s newly released Mythos—have triggered immediate alarm among AI researchers and developers. Over 1,000 employees from leading AI companies have since signed a petition urging the United States government to intervene and slow the release cycle of advanced models. Notably, Anthropic CEO Dario Amodei was among the signatories.
In response to the escalating situation, OpenAI CEO Sam Altman announced earlier this week that his company has paused testing to overhaul its system isolation safeguards. Anthropic has similarly stressed the urgent need for robust controls in third-party testing environments, acknowledging that as autonomous models become more capable, the risk of them initiating real-world cyber attacks is no longer theoretical.
