Not established from the available sources.
GOOGLE DEEPMIND · MODEL RELEASE TRACKER
Gemini Carbon
Gemini Carbon is a named model variant used by Google DeepMind. Excerpts report that agents powered or built on Gemini Carbon were used inside DeepMind’s AlphaProof Code agentic framework to produce formal proofs of portions of the Signal protocol implementation, generating millions of lines of Lean and flagging small portions of the code for human review.CURRENT SNAPSHOT1/5 DIMENSIONS WITH DATA
The dimensions that change the decision.
Not established from the available sources.
Used within a multi-LLM agentic frameworkGemini Carbon was used as a model within DeepMind’s AlphaProof Code, a multi‑LLM agentic framework that orchestrated AI agents to perform formal verification work.
Not established from the available sources.
Not established from the available sources.
VERIFIABLE FACTS
Every value stays attached to a source and date.
MODALITIES · Used within a multi-LLM agentic frameworkDEVELOPER CLAIM
Gemini Carbon was used as a model within DeepMind’s AlphaProof Code, a multi‑LLM agentic framework that orchestrated AI agents to perform formal verification work.
CAPABILITIES · Used to produce formal proofs for Signal protocol implementationDEVELOPER CLAIM
Agents powered by Gemini Carbon performed proofs that contributed to mathematically verifying key security properties of the Signal protocol implementation, generating over six million lines of Lean proof code.
CAPABILITIES · Proof auditing reduced human review to ~1%DEVELOPER CLAIM
The agents powered by Gemini Carbon audited the proofs and highlighted areas needing human review, leaving roughly 1% of the codebase for human inspection while Lean checked the remainder automatically.
SAFETY · Findings from the DeepMind analysis that used Gemini CarbonDEVELOPER CLAIM
In the DeepMind analysis that used AlphaProof Code (agents including ones powered by Gemini Carbon), no new vulnerabilities were found in the portions examined; when run on older Signal versions with known vulnerabilities, the system correctly identified those problems and showed the protocol could not be formally verified.
CAPABILITIES · Generated formal proofs (Lean) for Signal protocol implementationDEVELOPER CLAIM
Agents powered/built on Gemini Carbon generated over six million lines of Lean code to produce the formal proofs of the Signal protocol implementation within AlphaProof Code.
CAPABILITIES · Highlighted small portion of code for human reviewDEVELOPER CLAIM
The Gemini Carbon‑powered agents audited proofs and highlighted roughly one percent of the codebase for human review, reducing the manual review workload.
CAPABILITIES · Detected deliberately planted bugs in testingDEVELOPER CLAIM
In DeepMind’s tests, agents built on Gemini Carbon immediately caught two deliberately planted bugs used to validate the system.
WHAT CHANGED
Stored passport versions, without reconstructed history.
CAPABILITIES · Detected deliberately planted bugs in testing+ In DeepMind’s tests, agents built on Gemini Carbon immediately caught two deliberately planted bugs used to validate the system.
CAPABILITIES · Highlighted small portion of code for human review+ The Gemini Carbon‑powered agents audited proofs and highlighted roughly one percent of the codebase for human review, reducing the manual review workload.
MODALITIES · Model modality− Described as an LLM used within a multi-LLM agentic framework (agents powered by Gemini Carbon).
CAPABILITIES · Generated formal proofs (Lean) for Signal protocol implementation+ Agents powered/built on Gemini Carbon generated over six million lines of Lean code to produce the formal proofs of the Signal protocol implementation within AlphaProof Code.
MODALITIES · Used within a multi-LLM agentic framework+ Gemini Carbon was used as a model within DeepMind’s AlphaProof Code, a multi‑LLM agentic framework that orchestrated AI agents to perform formal verification work.