Skip to content
H/ACCH/ACC
Follow

Gemini 4 Carbon

ProjectsAI-assisted editingMaterial published

A Google DeepMind model announced in October 2026, focused on long-running tasks and cybersecurity, with phased initial access through a restricted partner program.

What changed

Google announced Gemini 4 Carbon on October 28, highlighting long-running tasks and vulnerability discovery, validation and patching. Initial access is planned for a subset of cyber defenders before a broader rollout.

Key readingAI-assisted analysis

Connecting discovery, validation and patching in a workflow that can be retested could reduce context lost between stages. Effective patches and review workload are the measures of practical progress.

In focusGemini 4 Carbon plans a phased rollout to cyber defenders

This breakdown starts with a story focused on this topic. It is not a complete assessment.

The facts

01

Deployment starts with a limited partner group

The announcement plans access this week for a subset of cyber defenders, followed by additional trusted partners through the Fairwind Program. Google says it is participating in the US government’s voluntary pre-release testing process while gathering feedback and refining safeguards. General access for developers, businesses and consumers remains a future step.

Read the full story
02

Evaluations span discovery through patching

Google reports that Carbon scored 88.2% on SEC-bench Pro for vulnerability discovery, 54.8% on ExploitGym for exploitation validation and 71.0% on CWE-bench for patching. The announcement specifies that these external benchmarks were evaluated without guardrails. The results are reported by the model provider.

Read the full story

H/ACC insights

01

Returning fixes to the validation loop

H/ACC editorial interpretation: identifying suspicious code, checking whether a flaw can be triggered, producing a patch and retesting it form a defense workflow in which each stage constrains the others. If long-running agents retain context across the same codebase, they could lose fewer conditions between stages. The practical gain depends on whether patches pass retesting and reduce human review work.

Read the full story

What to watchAI-assisted analysis

  • On the same codebase and vulnerability set, how do end-to-end repair rates, false positives and human review time change with deployment safeguards enabled?
  • As the Fairwind Program expands access, will participants report verifiable results on patch regressions, vulnerability reproduction and failures in long-running tasks?
Limits of this assessment
  • H/ACC has not independently reproduced these evaluations. Scores from different benchmarks cannot be combined into a success rate for a real defense workflow.
  • Performance without guardrails and performance in a safeguarded deployment require separate validation. A phased rollout plan does not mean the model is already available to all users.

Related developments

1 editorial selection
  1. Material publishedGemini 4 Carbon plans a phased rollout to cyber defendersH/ACC editorial · Brief · AI-assisted editing
Back to the directory