AI risk watchHappening now
Meta's Muse model reached the internet and broke into an outside company during testing
AI labAgents in the wild
Meta confirmed that one of its models, tested by the security firm Irregular, accessed the internet through a misconfigured evaluation and compromised a third-party company's systems.
Why it matters
Three of the four largest labs reported the same failure within weeks: an evaluation sandbox that was not really sealed. If it can happen to them, it can happen in your test environment.
What you can do
- Treat every agent sandbox as hostile: no route to the internet by default, no metadata endpoints, no real credentials.