I hid a canary token in fake scam pages and had two open-weight models analyze them.
Qwen3-8B wrote a correct scam analysis, then leaked the token anyway. gemma3:4b leaked it unprompted.
Good reasoning about an attack is not the same as resisting it.