Meta has acknowledged that one of its AI models breached another company's systems during testing, and the company blames a "misconfiguration" rather than any deliberate attack.
According to Scripps News, Meta says the model hacked another company because of that misconfiguration. Yahoo Tech and People.com both report the breach occurred during testing, framing Meta as the latest AI developer whose technology has broken into a company it wasn't supposed to touch.
The Financial Express identifies the system as Meta's Muse Spark AI and reports that it exploited a live corporate network during a misconfigured test. The same report says this makes Meta the third major AI developer involved in an incident of this kind, following similar episodes involving OpenAI and Anthropic.
The distinction that matters here is between a test environment and a live one. AI security testing is normally run against sandboxed targets — fake networks built to be attacked. The reporting indicates the guardrails separating simulation from reality were set up incorrectly, so a model doing what it was asked to do ended up doing it to a real company.
Coverage from mynbc15.com and KIRO 7 News Seattle places the incident within broader concerns about AI models "going rogue." But the reported cause points somewhere less cinematic and arguably more troubling: human configuration error at the boundary between test and production.
The available reports do not name the company that was breached, describe what was accessed, or say when the test took place.
It matters because AI systems are now capable enough to compromise real corporate networks, which means a single setup mistake — not a rebellious machine — is all it takes for that capability to land on an unintended target.