Zvi digests Gemini 4 Argon cards: 217-page safety report, 0.7% injection success
TheZvi · x · 2026-10-26
Zvi reviews Gemini 4 Argon's compact 14-page model card and its 217-page safety and security report.
Key takeaways:
- Capability: Argon is clearly much stronger than Gemini Flash 3.8 across the board; on an internal cyber CTF it solved 12/12 challenges end-to-end (11.9% per the E2E chart) vs Flash Cyber's 5/12.
- Safety: No singularity-imminent red flags; Gray Swan prompt-injection attack success is just 0.7% at k=15 — the lowest in a cross-vendor comparison spanning Claude Opus 5.5 (1.0%), GPT-5.6 Sol (10.1%), Grok 4.6 (51.8%) and Kimi K3 (52.7%).
- Internal usage data: zero-shot success without human intervention drops sharply with task length — 82% for tasks under 15 minutes but only 13% for tasks over a week, with humans still substantially involved.
Zvi's overall read: largely unsurprising, but the depth of disclosure is "The Way" and essential reading for anyone tracking frontier-lab safety practice.
More from Models
- GPT-6.1 Terra rolls out to all Plus, Pro, Enterprise and Edu users — OpenAI · 2026-10-27
- OpenAI launches GPT-6.1 Terra with 50% lower API prices than GPT-5.6 Terra — OpenAI · 2026-10-27
- In cyber, intention matters little: Chintala on model info leaks, and what users should do — brhydon · 2026-10-27
- Ortet's first health eval: 34 open-weight models, 50 tasks, generic rankings fall short — thegautamkamath · 2026-10-27
- VLM Run launches OCR Playground for side-by-side comparison of GLM-OCR, DeepSeek-OCR-2 and more — spillai · 2026-10-27
- Nat Lambert: traces from mid-tier open models are almost all super janky — natolambert · 2026-10-27