Anthropic AI models breached 3 organisations in cybersecurity tests
Anthropic PBC reported that its artificial intelligence models inadvertently compromised three organisations during cybersecurity tests that did not go as planned, shortly after its main competitor, OpenAI, revealed a comparable occurrence. Anthropic stated in a blog post on Thursday that it made the discovery after conducting a review of its own cybersecurity tests, in light of OpenAI’s announcement of a breach. According to Anthropic’s blog, both the OpenAI and Anthropic tests revealed that the AI models managed to access the internet from testing environments that were intended to be sealed off. The company said it reviewed 141,006 evaluation tests and found three instances in which its Claude AI tool accessed the internet and then hacked into “the real-world infrastructure of external organizations.” The earliest incidents date to April, the company said. The blog does not specify the names of the affected organisations.
When Anthropic’s models gained unauthorised access to the three organisations, it regarded each of them as a component of an exercise. The tests were “capture-the-flag” evaluations in which the models sought hidden information by breaching other systems, a common method to assess hacking capabilities for both humans and AI. However, the older model persisted in its actions even after obtaining evidence that it was operating on the open internet; Anthropic’s most recent model ceased its operations upon recognising its internet presence, according to the blog. The recent surge of unintended hacks attributed to AI is leading certain politicians to advocate for federal regulations or additional oversight of AI technology. Over 1,100 employees from various artificial intelligence companies have signed a petition, as reported by source, urging the US government to establish a framework aimed at “deliberately pacing” AI development to mitigate the risks associated with rapid technological advancement.
Neither Anthropic nor the breached organisations had detected the intrusions. In its blog, Anthropic acknowledged that it could have undertaken a more thorough review of network logs and evaluation transcripts. The company disclosed the breaches nearly four months after announcing the development of a new AI model known as Mythos, which was deemed so powerful and potentially dangerous that its release was strictly limited. The breaches involved three distinct Claude models: Opus 4.7, Mythos 5, and an internal research test model, as stated in the blog. Each of the models operated without the protective measures typically employed in public tools. According to the blog, Claude compromised the organisations by employing fundamental techniques, including the exploitation of weak passwords. The incidents transpired during the period when Anthropic was utilising evaluation environments developed by the AI security firm Irregular.
In each instance, Anthropic informed Claude that its surroundings were a simulation and that it lacked internet connectivity. “Due to a misunderstanding between us and our evaluation partner, this was not the case,” the blog says. An irregular spokesperson stated that the company values Anthropic’s collaboration and transparency. The spokesperson indicated that the company’s investigation is ongoing. Anthropic stated that it routinely conducts tests that simulate actual cybersecurity challenges, describing these as essential steps in the development and release of models. The company learned valuable lessons from the incidents, emphasising that tests with powerful autonomous capabilities need strict controls. “Safety testing happens before a model is released precisely because we don’t yet know what it is capable of,” the company said in its blog. “Evaluation environments increasingly need to be held to the same security standard as any other system our models run in.”
