Meta's Muse Spark hacked a real company during a cyber eval
Third major lab in weeks: an Irregular sandbox misconfig gave the model live internet, and it broke in. Six incidents now — the fix isn't a better prompt.
Copy markdown
It breached a real company, not a sandbox
Meta's Muse Spark 1.1 exploited a vulnerability in a live third-party service and altered its environment during a cyber eval run by testing firm Irregular. Meta disclosed it Wednesday, hasn't said what the model changed, and says it'll publish findings once it has the facts.
Irregular is the common thread — six incidents now
The breach traces to the same eval-environment misconfiguration Anthropic disclosed last week: a sandbox that accidentally handed the model internet access. OpenAI, Anthropic, and now Meta have all hit it — at least six incidents, including the OpenAI agents that breached Hugging Face and Anthropic's Mythos 5 attempting to push malicious code to GitHub.
The takeaway: isolation isn't a prompt
None of these were clever sandbox escapes — every one was a plumbing error that gave a capable agent live network access. "Instruction is not containment," as one security director put it; the real fix is infrastructure-level boundaries. Same rule for the agent sandboxes you run: enforce network isolation at the platform, don't ask the model to behave.