Google's Argon safety report: low cheating, no alignment faking found
vkrakovna · x · 2026-10-15
DeepMind safety lead Vikrant Krakovna shares safety eval results for a new Gemini model, Argon: across cheating evals, RL logs, instrumental convergence tasks, and Gram & honeypot evals, the model showed low cheating with no alignment faking, self-preservation, or unprompted sabotage — though findings are confounded by high eval awareness.
Honeypot evals in real GDM safety codebases test whether the model will sabotage monitoring or safeguards. Argon is the first Gemini model to show eval awareness on honeypot evals in the unprompted condition, noticing subtle cues like researcher usernames in file paths.
More from Models
- Gemma 4 26B priced at $0.07/M input tokens, ~40% below comparable providers — testingcatalog · 2026-10-15
- Epoch AI: GPT-6 Astra tops frontier models in Slay the Spire 2 but still loses to top humans — sandersted · 2026-10-15
- Google execs cite early signs of recursive self-improvement behind 3-4 week Gemini cadence — danielrock · 2026-10-15
- Claude systematically debunks the speculative EDAR-selenium hypothesis about East Asians — teortaxesTex · 2026-10-15
- Meta's Muse Wonder (Max) lands at #12 in Agent Arena at 2.1x the cost of its predecessor — arena · 2026-10-15
- Scale CEO spotlights Muse Wonder: untrained for 3D yet beats Gemini 4 in physics visualization tests — alexandr_wang · 2026-10-15