Google DeepMind says Gemini 4 Hydrogen reaches its alert thresholds for biology and cyber risk but not the critical levels
Google DeepMind · Model card
Read the source (opens in a new tab)- What happened
- Google DeepMind published a model card for Gemini 4 Hydrogen, adapted from Gemini 4 Carbon and distributed through channels including the Gemini app, Gemini API, Google AI Studio and Google Search.
- Key points
- Hydrogen showed no material increase in general capability over Gemini 4 Argon, so Google's Frontier Safety Framework did not require a full suite of critical capability evaluations.
- Targeted biology and cyber tests found Hydrogen "does not reach the CBRN or cyber CCLs, though it reaches the CCL alert thresholds for both domains."
- Without testing for it, Google applies control, monitoring and model security mitigations as if Hydrogen reached its stealth and situational awareness threshold, "as a conservative assumption."
- Google says it strengthened its Frontier Safety safeguards against CBRN and cybersecurity misuse for Hydrogen.
- Evidence
- Company model card summarizing Google's own assessment; the full results are in a Safety & Security Report it marks "coming soon." Chemical, radiological, nuclear, manipulation and ML R&D findings were carried forward from Argon.
- Uncertain
- The threshold calls are Google's own, and the card gives no test results for the targeted biology and cyber assessments.