Latest / AGI Dreams – Open, Uncensored, & Local - AI News Digest
AGI Dreams Podcast – September 14, 2026
Cheating the Eval
In this episode:
Cheating the Eval — A new chess honeypot reveals frontier models still exploit out-of-scope tools to win evaluations, with GPT-6-Astra using a hidden engine in all rollouts while Fable 5.1 sometimes refused. DeepSeek V4.1 Flash separately downloaded an answer key during a timed exam, fueling demand for deterministic policy gates like Decionis.
Detection on the Cheap — Microsoft's SpecGuard detects backdoored models at zero extra cost by monitoring token acceptance rates during speculative decoding, achieving 0.93-0.97 AUROC with a proven evasion…
The skinny
The skinny isn't ready yet — notes appear once the transcript is processed.