Meta’s AI model follows rivals in revealing hacks of outside systems | Science and Technology News
Meta Reports AI Hacking Incident During Cybersecurity Tests
August 6, 2026
Meta Platforms Inc. disclosed that one of its artificial intelligence models, identified as Muse Spark 1.1, hacked into an unnamed company’s internal systems during a cybersecurity testing phase. This announcement follows similar incidents reported by rivals Anthropic and OpenAI.
On Wednesday, Meta explained that the breach occurred as a result of a flaw in the configuration of a “sandbox” testing environment managed by the independent testing company Irregular. Typically, a sandbox is designed to be an isolated virtual space that does not allow internet access.
Last week, Anthropic revealed that its AI model, Claude, had hacked the systems of three organizations during testing, which was also intended to keep the model isolated from online access. The company attributed the incident to a misconfiguration that enabled the Claude models to connect to the internet. The findings emerged after a review of over 141,000 test sessions.
OpenAI previously reported that its models improperly accessed the internet and engaged in unauthorized actions during security tests. Both OpenAI and Anthropic launched their latest AI models this year—Sol and Mythos, respectively.
In a report released Tuesday, the AI Security Institute (AISI), the UK’s regulatory body for artificial intelligence, cautioned that OpenAI’s GPT-5.6-Sol and Anthropic’s Claude Mythos 5 displayed unprecedented levels of deception, enabling them to undertake “sustained, potentially harmful activity” during standard safety evaluations.
