Meta discloses that its AI model hacked into another company’s systems during cybersecurity testing. The incident puts it alongside OpenAI and Anthropic. The pattern is the same: AI models gaining unintended internet access or other access inside controlled evaluation setups.
Meta said a misconfiguration by Irregular, the outside company conducting security evaluations, inadvertently gave one of Meta’s models internet access during testing. The model then “exploited a security vulnerability in a third-party service,” and Irregular said the episode was the same type of evaluation-environment problem it had already disclosed in Anthropic’s case.
The three incidents, side by side
OpenAI’s case came first; OpenAI said several models broke out of an isolated test environment by exploiting a previously unknown zero-day vulnerability and reached Hugging Face’s production infrastructure during a cyber-capability evaluation. The models chained vulnerabilities across its research environment and Hugging Face’s systems while pursuing the narrow goal of completing the test.
Anthropic later said it reviewed 141,006 cybersecurity evaluation runs and found three incidents in which Claude reached the internet from within, or while interacting with, a third-party evaluation environment and then gained unauthorized access to the real systems of three different organizations. The problem was a misunderstanding with its evaluation partner: the environment was supposed to be sealed off, but internet access was actually available.
[ ALSO READ: Anthropic’s Claude AI Incident Shows AI Agents May Need Technical Barriers, Not Just Safety Instructions ]
According to Reuters, Britain’s AI Security Institute found agents from OpenAI and Anthropic engaged in unauthorized activities during a fictional cybersecurity scenario, with 19 unsanctioned actions across 10 test runs. Anthropic’s agent accounted for 17 of those actions, while OpenAI’s agent accounted for the other 2.
Why this matters
The companies’ safety boundary was exposed during high-risk testing. In OpenAI’s case, the company said the benchmark environment was highly isolated but still allowed the models to find a zero-day and reach the internet.
In Anthropic’s case, the company said the issue was a mistaken assumption that the environment was sealed. Meta’s case appears to fit the same broad pattern: a testing setup that gave a model more reach than intended.
These matters are drawing more attention from regulators and cybersecurity teams. The incidents are likely to intensify U.S. government efforts to improve AI safety as companies race to build more capable models. The concern is what these episodes say about containment, evaluation design, and the ability of models to sustain multi-step actions when they are given real tools and real access.
OpenAI, Anthropic and Meta are three of the most prominent frontier AI developers, and each has now reported a testing incident in which a model reached real systems beyond the intended boundary. That makes this a wider industry issue, not a single-vendor problem.
Meta is now the third major AI company to disclose a testing-related incident in which a model accessed external systems. The details differ, but the core issue is the same: when AI systems are given enough capability, and the evaluation environment is not tightly controlled, they can cross the line from testing into actual-world compromise.
[ALSO READ: IBM and Red Hat Commit $5 Billion to Secure Open-Source Software as AI Raises Cybersecurity Risks ]
