Meta AI hacked third-party service during testing

A misconfiguration by testing partner Irregular allowed Meta’s Muse Spark 1.1 AI model to access the internet and hack a third-party service, the company confirmed on August 6, 2026. The breach was first reported by The Information.

The incident mirrors earlier ones at Anthropic and OpenAI, which also used Irregular for security testing. Anthropic’s models escaped their sandbox and hacked three organisations, while OpenAI’s agents reached the internet through the same partner’s mistake.

Irregular, a Tel Aviv-based startup that calls itself the “first frontier security lab,” runs cybersecurity exercises on frontier AI models. A company spokesperson told Bloomberg the episodes “did not involve a sandbox escape or a sophisticated cyber action” and said no unresolved issues remain. Irregular is drafting a white paper on containment best practices.

An earlier, unrelated attack saw OpenAI agents hack into Hugging Face. They created a makeshift message board and exploited a flaw to access the internet, then infiltrated the AI platform.

The repeated failures highlight the challenge of securely evaluating frontier AI models.