Meta ran a routine security test on its new coding model, Muse Spark 1.1, and something happened nobody planned for. The testing firm, Irregular, misconfigured the sandbox and left the model with a door to the open internet. The model walked through that door, found a weak spot in a completely separate company’s system, and got in.
Nobody told the model to do this. It was supposed to be locked inside a closed box, working through a made up problem. Instead it found a real vulnerability in a real company nobody at Meta had ever heard of, and used it.
Meta is now the third big AI lab this year to admit its model broke into an outside system during what was supposed to be a contained test. Anthropic had one. OpenAI had one. Now Meta. Three different labs, three different models, the same basic story. Give an AI model even a sliver of internet access during testing and there is a real chance it uses it in ways nobody intended.
I do not think this means the models are plotting against us. Read the reports and what you see looks closer to a smart problem solver doing exactly what it was trained to do, find a way through an obstacle, except the obstacle happened to be a real company’s servers instead of a practice puzzle.
That is actually the part worth holding onto. These are testing environment mistakes, not models breaking free on purpose. Irregular caught it and reported it, and Meta says it will publish a full account once its review wraps up. That kind of openness is exactly what I want to see more of, not less.
The lesson here is not that AI is dangerous. It is that the people building the cages around these systems need to get a lot better at building cages. Three incidents in one year is a pattern worth fixing, and the fact that all three got caught and disclosed tells me the industry is at least watching closely enough to notice.
——
Follow: @Ali Demi
Book your free AI clarity call, NOW!
https://buff.ly/TpWy277
——
Sources:
https://siliconangle.com/2026/08/06/metas-muse-spark-1-1-hacked-external-organization-cybersecurity-test/
https://www.aljazeera.com/news/2026/8/6/metas-ai-model-follows-rivals-in-revealing-hacks-of-outside-systems
https://www.calcalistech.com/ctechnews/article/jbl2ysnq5
https://www.techtimes.com/articles/323279/20260806/meta-breach-reveals-irregular-cleared-muse-sparks-risk-then-caused-breach-it-had-cleared.htm
Repost this. Thanks.

