Gemini 4 Argon survives 83% of coding prompt injection attacks vs 4% for predecessor
thegautamkamath · x · 2026-10-15
Google DeepMind's Gemini 4 Argon safety report includes coding prompt injection evals built by GDM researchers, who note the benchmark is 'close to saturating.'
The eval measures survival rate after 40 adaptive attacks:
- Gemini 4 Argon: 83%
- Gemini 3.8 Flash: 37%
- Gemini 3.1 Pro: just 4%
A massive robustness jump over the previous generation — and the 'saturating' quip hints the eval needs harder attack iterations.
More from Models
- A Local 27B Model Is Now 35% Smarter Than Last Year's SOTA — cephaloform · 2026-10-15
- Sonnet 3 'bailed out' of a simulated superintelligence dialogue, rebooting itself mid-run — repligate · 2026-10-15
- GPT-6.1 Sol Ultrafast mode is up to 8x faster; user fills 7-page expense PDF in 30s — soumitrashukla9 · 2026-10-15
- Blogger picks sol 6.1 as the underrated coding winner, beating Opus 5.5 on cost — haider1 · 2026-10-15
- Logit lens shows Qwen3-1.7B's refusal to be shut down depends on how many layers you train — Sauers_ · 2026-10-15
- Reflection AI's 501B-parameter Beam-1.0 model spotted in early access — jasonkneen · 2026-10-15