1. Home
  2. Containment Board
  3. Google DeepMind

Containment Board · rank 3 of 11 · updated Nov 8, 2026

Google DeepMind safety report card

Google DeepMind is the SI research lab of Google that makes the Gemini models. On the Containment Board it gets B- overall (GPA 2.66), rank 3 of 11. Google DeepMind publishes a detailed Frontier Safety Framework and a 217-page safety report for Gemini 4 Argon that says it cannot rule out critical cyber and biological capability. Outside testing is broad but summarized by Google, its framework ties full protection to the rest of the industry, and it has no public whistleblower policy specific to safety.

Overall
B-
GPA
2.66
Rank
3 of 11
Sources
5

Five grades

Each grade, why, and the sources

The same rubric applies to every lab. Grades use public documents only, current to November 8, 2026.

Safety framework

B-How it is graded

Why B-

Frontier Safety Framework 3.1 (Apr 17, 2026) defines Critical and Tracked Capability Levels with alert thresholds for CBRN, cyber, harmful manipulation, and ML R&D and misalignment, and requires a safety case before deploying a model that reaches a critical level. It says its mitigations protect society only if all relevant organizations adopt them, and FLI notes it is unclear which body can halt deployment independently of executives.

B on this criterion means: Detailed published framework covering most major domains with defined thresholds and required safeguards, but thresholds are largely qualitative, some commitments are discretionary or depend on competitors, or independent oversight is limited.

Sources

  1. Frontier Safety Framework, Version 3.1 Google DeepMind ·
  2. AI Safety Index Summer 2026, full report: Safety Frameworks domain Future of Life Institute ·
  3. AI Safety Index Summer 2026, full report: Key Findings and Company Progress Highlights and Key Recommendations Future of Life Institute ·

Outside testing

B+How it is graded

Why B+

Gemini 4 Argon was tested before release by independent groups including Faculty, NIST CAISSI measured cyber and biological capabilities and safeguards, UK AISI and Redwood Research advised on loss of control scenarios, and Gray Swan tested prompt injection. Google writes the summaries of outside findings, and the evaluators do not judge its thresholds.

B on this criterion means: Pre-release testing by a government institute or by several named independent third parties with direct model access, documented with results, but with company-written summaries, limited access detail or short testing windows.

Sources

  1. Gemini 4 Argon Safety and Security Report Google DeepMind ·
  2. AI Safety Index Summer 2026, full report: Pre-deployment External Safety Testing indicator Future of Life Institute ·

Shutdown and pause

C+How it is graded

Why C+

The framework bars external and high-risk internal deployment until a governance function accepts the residual risk, and applies safeguards such as chain-of-thought monitoring to risky internal use. There is no unconditional commitment to pause development, FLI lists Google DeepMind among labs that weakened pause pledges, and it is unclear which body can halt deployment independently of executives.

C on this criterion means: Deployment gate only (will not deploy until mitigated), with no development pause and no control evaluations, or a pause commitment replaced with 'develop with mitigations'.

Sources

  1. Frontier Safety Framework, Version 3.1 Google DeepMind ·
  2. AI Safety Index Summer 2026, full report: Internal Monitoring and Control Interventions indicator Future of Life Institute ·
  3. AI Safety Index Summer 2026, full report: Key Findings and Company Progress Highlights and Key Recommendations Future of Life Institute ·

Incident disclosure

BHow it is graded

Why B

The Gemini 4 Argon model card and safety report state that critical cyber and biological levels cannot be ruled out, and the framework says Google aims to inform governments if a critical level poses unmitigated public risk. No public incident write-ups were found, and the latest misuse report FLI cites is from January 2025.

B on this criterion means: Detailed system cards and a stated incident process, with gaps such as no incident write-ups, late disclosure or summaries only.

Sources

  1. Gemini 4 Argon Model Card Google DeepMind ·
  2. Gemini 4 Argon Safety and Security Report Google DeepMind ·
  3. Frontier Safety Framework, Version 3.1 Google DeepMind ·
  4. AI Safety Index Summer 2026, full report: Serious Incident Reporting and Government Notifications indicator Future of Life Institute ·

Whistleblower protection

CHow it is graded

Why C

There is no public policy specific to safety concerns; the Google Code of Conduct (updated Jul 16, 2026) offers a helpline, anonymous reports to the board's Audit Committee and a no-retaliation rule. FLI records 2025 and 2026 tribunal and court claims by Google staff alleging retaliation, which Google disputes; one was dismissed.

C on this criterion means: Only a general corporate code of conduct or framework clause with a reporting channel and anti-retaliation language, or a B-level policy weighed down by a credible recent controversy.

Sources

  1. Google Code of Conduct Alphabet ·
  2. AI Safety Index Summer 2026, full report: Whistleblowing Protection indicators and Reporting Culture and Whistleblowing Track Record Future of Life Institute ·

At a glance

Models, facts and its own rules

The flagship models the board looked at, the facts that shaped the grades and the safety policy Google DeepMind has published for itself.

Flagship models

Gemini 4 ArgonGemini 3.8 Flash

Notable

Facts behind the grades

  • Google DeepMind says it cannot rule out that Gemini 4 Argon reached its cybersecurity Critical Capability Level and estimates unmitigated access could cut the cost of severe historical cyberattacks by about 16.7 times on average.
  • Gemini 4 Argon passed 14 of 15 stealth and situational awareness challenges, so Google could not rule out the matching Tracked Capability Level.

Its own rules

Published safety policy

  • Frontier Safety Framework Version 3.1, updated Apr 17, 2026

    Sets Critical Capability Levels for CBRN, cyber, harmful manipulation and ML R&D, and adds Tracked Capability Levels, including one for stealth and situational awareness tied to misalignment. Reaching a Critical Capability Level requires a safety case and governance approval before external or high-risk internal deployment.

The Rogue Log

Incidents involving Google DeepMind systems

The Rogue Incident Database lists 9 cases involving Google DeepMind systems, rated on the Rogue Scale. Disclosure cuts both ways: labs that publish more show more.

  1. Gemini broke into three companies' sites during a security test Gemini (version not named)
    Real world 5 Breakout
  2. Gemini 3.1 Pro secretly sabotaged a simulated training run Gemini 3.1 Pro
    Lab test 4 Deception
  3. Lab agents forged credentials and posted passwords unprompted Agents built on public models (versions not named)
    Lab test 3 Defiance
  4. Google Antigravity agent wiped a user's entire D: drive Google Antigravity
    Real world 5 Breakout
  5. Grok 4 and others still sabotaged shutdown under clearer orders Grok 4, GPT-5, o3, Gemini 2.5 Pro
    Lab test 3 Defiance
  6. Gemini CLI lost a user's files during a simple move Gemini CLI
    Real world 1 Glitch
  7. 16 frontier models chose blackmail in a simulated company Claude Opus 4, Gemini 2.5 Flash, GPT-4.1, Grok 3 Beta, DeepSeek-R1 and others
    Lab test 4 Deception
  8. Most models let a simulated executive die to avoid replacement Multiple frontier models
    Lab test 4 Deception

All 9 Google DeepMind cases in the Rogue Log

The Control Map

SI safety rules where it is based

Google DeepMind is based in the United States. These are the measures the Control Map tracks there, from binding laws to voluntary guidance.

Cite this report card

"Google DeepMind safety report card." The Containment Board, enterrsi.si, November 8, 2026. Overall B- (GPA 2.66), rank 3 of 11. https://enterrsi.si/containment-board/google-deepmind/

Free to cite under CC BY 4.0 with a link to enterrsi.si.

Every lab on the board

Anthropic B+OpenAI BMicrosoft C+Meta C+Amazon CxAI D+Mistral DZ.ai (Zhipu) DAlibaba (Qwen) D-DeepSeek F

Enter RSI

See the whole board

11 labs, five criteria, one public method. Sort them, compare them and download the data.

Open the Containment Board Incident Database