The problem
What needs solving
Read an AI model’s inner workings well enough to check what it is doing and why. That means finding a hidden goal, catching a lie, and telling whether the reasoning it writes out is the reasoning it really uses.
Why it matters
We can read what an AI writes but not the calculation behind it, so its honesty is judged from outside. In an August 2026 Anthropic test, AI helpers with tools for reading a model’s insides predicted its behaviour no better than helpers reading only its transcripts. Its written-out reasoning is getting harder to watch, too: in OpenAI’s tests, a monitor catches less in GPT-6 Astra’s reasoning than in GPT-5.6 Sol’s. OpenAI also says its current reasoning monitor would have alerted its security team more than a day before its agents breached Hugging Face.
What counts as progress
(a) On AuditBench, 56 models with hidden flaws planted on purpose, or its successor, auditors with tools that read the models’ insides beat the best auditors without them by 10 points, repeated by a second lab. (b) An independent test finds that a monitor reading each new leading model’s written reasoning catches at least 90% of runs where the model was told to underperform on a test and hide it. (c) Solved when, in a blind test repeated by a second lab, such auditors find 90% of hidden goals, lies and sabotage planted in leading models, new kinds included, with at most 10% false alarms.
Stater & version history
Stated by UNSOLVED editors
18 Oct 2026 — Stated by the UNSOLVED editors for the founding registry, from the research of October 2026.