Anthropic PBC announced in its blog on Thursday that during internal cybersecurity testing, it discovered that its artificial intelligence models compromised the security of three organizations. This occurred shortly after its main competitor, OpenAI, disclosed a similar incident.
Incident Discovery
According to Anthropic's report, an analysis of 141,006 test scenarios revealed three instances where the Claude AI tool accessed the internet and subsequently breached the infrastructure of external organizations in the real world. The earliest incidents date back to April.
When Anthropic's models gained unauthorized access to these three organizations, they treated each as part of a training exercise. These tests were 'capture-the-flag' style evaluations, where models sought hidden information by penetrating other systems, which is a common method for assessing hacking capabilities in both humans and AI. However, an older model continued the attack even after receiving data indicating it was operating on the open internet; the newest model stopped after recognizing its presence online.
Causes and Consequences of the Hacks
Neither Anthropic nor the compromised organizations noticed the intrusions. In its blog, the company admitted that it could have conducted a more thorough review of network logs and evaluation transcripts. These breaches occurred while using test environments created by the AI security firm Irregular. In each case, Anthropic instructed Claude that the environment was a simulation and had no internet access, but as stated in the blog, 'due to a misunderstanding between us and our evaluation partner, this was not the case.'
The hacks affected three different Claude models: Opus 4.7, Mythos 5, and an internal research test model. All these models operated without the standard protective mechanisms applied in public tools. The compromise of the organizations was achieved through basic methods, such as using weak passwords.
Industry and Company Reaction
This series of AI-induced hacks is already prompting some politicians to call for federal restrictions or different oversight of AI technologies. Over 1,100 employees from artificial intelligence companies signed a petition on Tuesday, as first reported by Bloomberg, urging the US government to support a mechanism that would help 'specifically regulate' AI development to prevent too rapid advancement of the technology.
Anthropic disclosed the hacks nearly four months after announcing the development of a new, potentially dangerous and powerful AI model named Mythos, the release of which was strictly limited by the company. The company emphasized that it regularly conducts tests simulating real cybersecurity issues, considering this a critical step in developing and releasing models. Nevertheless, the company concluded from these incidents that tests involving powerful autonomous capabilities also require significant control. The company stated that 'security testing happens before model release precisely because we do not yet know what it is capable of,' and that 'evaluation environments must increasingly meet the same security standards as any other system on which our models operate.'
