On Tuesday, Google announced a broader availability of its Gemini 4 Carbon, emphasizing the role of AI agents in enhancing security verification of the Signal protocol’s code. This move signifies a major step forward in AI-driven cybersecurity solutions.
Expanded Access to Developers
Developers and businesses that subscribe to Google AI Ultra can now leverage Carbon through the Gemini API and Gemini for business. An introductory pricing of $2/$10 per million input/output tokens is in place until November 30, after which prices will increase. Previously, access was limited to select members of Google’s Fairwind Program, who continue to require membership to utilize Carbon for cyber applications.
Google highlighted that Carbon has already been instrumental in identifying zero-day vulnerabilities, as demonstrated by its use by Wiz in two major SaaS environments. The Cloud Vulnerability Research team at Google utilized Carbon to confirm the exploitability of two bugs in a widely used open-source project, similar in severity to the notorious PHP2Shell vulnerabilities.
AI and Formal Verification
In a separate development, Google DeepMind has introduced AlphaProof Code, which combines human expertise with AI agents to formally verify software against defined requirements. This was applied to libsignal, the reference implementation of the Signal protocol, which is also used in WhatsApp and Google Messages.
AI agents, operating on Carbon and Claude Opus 5.5, collaborated with human experts to draft specifications, generating approximately 6.5 million lines of Lean proofs. These proofs uphold key security features such as forward secrecy and authentication, with Lean rejecting any invalid lines. This collaboration achieved in four weeks what traditionally took years, showcasing the potential of AI in formal verification.
Implications for Future Security
The study demonstrated the efficacy of AI in detecting real-world security flaws by reintroducing previously patched libsignal bugs, which disrupted security guarantees. However, this analysis was confined to Google’s build of the code, excluding Signal’s own applications and certain side channels.
The findings suggest that AI can make formal verification feasible for real-world systems, a process traditionally considered too costly. As AI-driven security threats continue to evolve, the rigorous verification of the Signal protocol enhances confidence in the privacy assurances of the code, according to Signal CTO Ehren Kret.
Google is making the specifications and proofs publicly available, contributing to the broader cybersecurity community. This development marks a significant milestone in AI’s role in advancing software security.
