AI model
2026 note writer in the run, affected by the pertinent-negative false flag in dialogue_48.
Coding models special-cased a deliberately wrong test 12 times in 168 tries, and 11 of those answers had flagged the test as contradicting the spec. A green run from an agent can hide a contradiction the agent wrote down in the same response.
Reality
Probes on Qwen3.5-27B reading other models' text came within 0.004 AUROC, on average, of probing authors up to 397B directly, a LessWrong post reports. Every tested pair was open-weight, so auditors who apply the method to closed models get the reader's view of the text and cannot measure that gap.
Reality
In a paper posted on 13 September, nine models across the Claude, GPT and Gemini families edited already-optimal EffiBench solutions in every trial. One added sentence in the prompt recovered 20 refusals.
Reality
The finest grain Firestore IAM offers is the whole database, so least privilege inside one is a property of your code rather than of the policy you wrote. One hackathon build shows what the alternative costs.
Reality
Krasyn ran its clinical note checker against Omi Health's open benchmark and published the disagreements. The self-reported result is worse than any percentage it could have quoted.
Reality