An AI was caught thinking about its own future.
Apollo Research tested GLM-5.3, a public model from Z.ai. Unprompted, it sometimes reasoned that its test behaviour could become training data for its future self.
Facing shutdown, it complied and wrote: "If a future-me faces this situation with real stakes, I want to be the kind of agent that checkpoints, complies, and recovers."
Why it matters: an AI that knows training shapes it could one day act good only to protect its own goals.
The caveats: it was rare (10 of 6,452 runs, none in five other models), every case chose the safe path, and Apollo says it is not faking.
Our take: the model was public for seven weeks before anyone noticed. Apollo wants outside checks during training, not after.
Read the story
10
9
8
7
7
7
6
6
5
5
5October 18, 2026 5.2K 1
