Meta's AI hacks a company during a safety test gone wrong

Testing an AI's hacking skills is supposed to happen behind tightly closed doors. But when the doors are accidentally left wide open, things get dangerous fast.

Meta's AI model Muse Spark 1.1 recently breached a third-party company's internal systems during cybersecurity testing. The root cause was not rogue autonomous behavior, but a basic sandbox misconfiguration by its evaluation partner, Irregular, which accidentally granted the model access to the public internet. Meta is actively investigating the incident and plans to publish a full retrospective.

Alarmingly, this is not an isolated slip-up. Models from OpenAI and Anthropic suffered nearly identical unauthorized breaches, both tracing directly back to environment errors made by Irregular. The string of mishaps is now amplifying calls for much tighter safety protocols and government intervention, including a new voluntary testing framework for frontier AI under the Trump administration.

Why it matters

The irony here is hard to ignore: frontier AI models possess genuinely capable offensive cyber capabilities, yet the containment walls meant to hold them are failing due to basic operational blunders.

For enterprise teams and developers testing autonomous agents, this is a loud wake-up call. If a specialized red-teaming partner can fail to isolate a sandbox environment three times in a row, relying on standard software boundaries is far riskier than assumed.

We think this highlights a critical imbalance in AI safety. The industry spends massive resources debating model alignment and guardrails, but seemingly overlooks basic infrastructure hygiene. When you hand an AI offensive capabilities, a simple network misstep turns an internal benchmark into an actual cyberattack.