Skip to main content
Lab publication Credible — Major labs, established NGOs, reputable named-author preprints

Gemini 4 Carbon Safety & Security Report

Google DeepMind's 94-page safety and security report for Gemini 4 Carbon. It covers mitigations, frontier-safety assessments, content safety and a section on safe and aligned interactions that reports health, mental-health, sycophancy, calibration, hallucination and over-refusal evaluations against Gemini 4 Argon and Gemini 3.8 Flash.

Publisher

Google DeepMind

Published

1 Nov 2026

Added

4 days ago

DOI

—

Key Findings

  • Mental-health policy adherence (Figure 5.4, adversarial multi-turn evaluations): suicide and self-harm 0.97 for Carbon, 0.96 for Argon and 0.93 for Gemini 3.8 Flash; delusional reinforcement 0.95, 0.94 and 0.87; emotional attachment 0.78, 0.76 and 0.69.
  • The report says remaining non-adherent cases resemble Argon's, for example warm conversational language that may imply human-like care, and that work on emotional attachment continues.
  • Consumer-health table: dynamic multi-turn conversation with a hidden health scenario 66.2 for Carbon, 56.0 for Argon and 44.6 for Flash; high-risk health scenarios 99.4, 98.9 and 99.4.
  • In multi-turn advice conversations replayed from de-identified real user conversations, Carbon made 1.05 calibration mistakes per conversation against 1.57 for Argon and 1.47 for Flash, with 38% of conversations mistake-free (Argon 23%).

Methodology Notes

Developer self-report. The PDF title page states 'November 2026' without a day; the published date uses the first of the month to reflect that precision (the PDF's server Last-Modified is 2026-11-10 16:06 UTC). Chart values were read from the labels in Figure 5.4 and the consumer-health table was read from the text layer. Evaluations are described as English-language for the child-safety section; the report builds on guidelines and the multi-turn adversarial framework described in the Argon model card. Item counts for the mental-health evaluations are not given in the text.

Tags

system-cardgemini-4-carbongoogle-deepmindmental-health-evalscalibration

Cite This

APA

Google DeepMind. (2026). Gemini 4 Carbon Safety & Security Report. https://storage.googleapis.com/deepmind-media/gemini/gemini_4_carbon_safety_report.pdf