How Researchers Used Gemini Carbon to Prove Signal's Encryption in Lean

Google's formal verification sprint shows how frontier models may accelerate trustworthy cryptography—and what limits remain.

6 min read

Google's release of Gemini 4 Carbon to paying developers was already a benchmark story; the science headline was what researchers did with it next. Teams at Google used Carbon to generate millions of lines of Lean—a proof assistant language where every step is checkable—to reason about Signal's encryption implementation. They targeted formal guarantees around core security properties that human expert teams historically proved over years.

Why formal methods matter now

Machine learning skeptics often ask whether stochastic models belong anywhere near cryptography. The right framing is complementary division of labor: models propose proof sketches and tactic scripts; proof assistants reject anything that does not compile in the logic. The output is not "trust the LLM"; the output is a certificate human reviewers can audit.

Signal's codebase is a cultural touchstone for privacy engineers. Demonstrating that automated proof search can cover meaningful lemmas there is a morale boost for engineers tired of AI hype slides. It also pressures other messengers and VPN stacks to publish what has and has not been verified.

AlphaGenome 2 in the same week

Google parallel-launched AlphaGenome 2, a DNA-sequence model aimed at predicting how variants affect gene regulation in specific cell types. The system reportedly surfaces nearly twice as many candidate variants tied to expression changes compared with its predecessor—still research-only, but relevant to precision medicine pipelines.

Pairing formal software verification with biotech foundation models in one news cycle illustrates how "science" AI is diverging from chat UX. JSIPE readers should notice the shared infrastructure: gigantic context windows, curated evaluation sets, and ruthless human gatekeeping before anything touches patients or production TLS stacks.

Limits and caveats

Lean proofs about a specific snapshot of Signal do not immunize the messenger's entire attack surface—key exchange policies, client update paths, and server-side metadata policies remain outside any single formal artifact. Likewise, genome models inherit dataset bias; doubling discovered variants is not doubling clinical validity.

What labs should copy

If you run a security-critical code path, pilot a formal-methods sandwich: LLM drafting, proof assistant filtering, human maintainer sign-off. Budget for maintainer time; proofs rot when code drifts. If you run computational biology, treat AlphaGenome-class tools as hypothesis generators feeding wet-lab confirmation—not as diagnostic products.

The through-line for November 2026 is disciplined acceleration: machines explore vast search spaces; institutions keep the bar for deployment unchanged or higher.

Operational teams should treat vendor announcements as triggers for regression tests, not as finish lines.

Document who approved production changes, which prompts were pinned, and what rollback looked like when the first anomaly appeared.

Security and compliance stakeholders need plain-language briefings: what data leaves the perimeter, what logs exist, and what happens on failure.

Leaders who invest in evaluation harnesses this quarter usually spend less on firefighting next quarter—especially when defaults shift underfoot.

Operational teams should treat vendor announcements as triggers for regression tests, not as finish lines.

Document who approved production changes, which prompts were pinned, and what rollback looked like when the first anomaly appeared.

Security and compliance stakeholders need plain-language briefings: what data leaves the perimeter, what logs exist, and what happens on failure.

Leaders who invest in evaluation harnesses this quarter usually spend less on firefighting next quarter—especially when defaults shift underfoot.

Operational teams should treat vendor announcements as triggers for regression tests, not as finish lines.

Document who approved production changes, which prompts were pinned, and what rollback looked like when the first anomaly appeared.

Security and compliance stakeholders need plain-language briefings: what data leaves the perimeter, what logs exist, and what happens on failure.

Leaders who invest in evaluation harnesses this quarter usually spend less on firefighting next quarter—especially when defaults shift underfoot.

Operational teams should treat vendor announcements as triggers for regression tests, not as finish lines.

Document who approved production changes, which prompts were pinned, and what rollback looked like when the first anomaly appeared.

Security and compliance stakeholders need plain-language briefings: what data leaves the perimeter, what logs exist, and what happens on failure.

Leaders who invest in evaluation harnesses this quarter usually spend less on firefighting next quarter—especially when defaults shift underfoot.

Operational teams should treat vendor announcements as triggers for regression tests, not as finish lines.

Document who approved production changes, which prompts were pinned, and what rollback looked like when the first anomaly appeared.

Security and compliance stakeholders need plain-language briefings: what data leaves the perimeter, what logs exist, and what happens on failure.

Leaders who invest in evaluation harnesses this quarter usually spend less on firefighting next quarter—especially when defaults shift underfoot.

Operational teams should treat vendor announcements as triggers for regression tests, not as finish lines.

Document who approved production changes, which prompts were pinned, and what rollback looked like when the first anomaly appeared.

Security and compliance stakeholders need plain-language briefings: what data leaves the perimeter, what logs exist, and what happens on failure.

Leaders who invest in evaluation harnesses this quarter usually spend less on firefighting next quarter—especially when defaults shift underfoot.

Operational teams should treat vendor announcements as triggers for regression tests, not as finish lines.

Document who approved production changes, which prompts were pinned, and what rollback looked like when the first anomaly appeared.

Security and compliance stakeholders need plain-language briefings: what data leaves the perimeter, what logs exist, and what happens on failure.

Leaders who invest in evaluation harnesses this quarter usually spend less on firefighting next quarter—especially when defaults shift underfoot.

Operational teams should treat vendor announcements as triggers for regression tests, not as finish lines.

Document who approved production changes, which prompts were pinned, and what rollback looked like when the first anomaly appeared.

Security and compliance stakeholders need plain-language briefings: what data leaves the perimeter, what logs exist, and what happens on failure.

Leaders who invest in evaluation harnesses this quarter usually spend less on firefighting next quarter—especially when defaults shift underfoot.

Operational teams should treat vendor announcements as triggers for regression tests, not as finish lines.

Document who approved production changes, which prompts were pinned, and what rollback looked like when the first anomaly appeared.

Security and compliance stakeholders need plain-language briefings: what data leaves the perimeter, what logs exist, and what happens on failure.

Leaders who invest in evaluation harnesses this quarter usually spend less on firefighting next quarter—especially when defaults shift underfoot.

Operational teams should treat vendor announcements as triggers for regression tests, not as finish lines.

Document who approved production changes, which prompts were pinned, and what rollback looked like when the first anomaly appeared.

Security and compliance stakeholders need plain-language briefings: what data leaves the perimeter, what logs exist, and what happens on failure.

Leaders who invest in evaluation harnesses this quarter usually spend less on firefighting next quarter—especially when defaults shift underfoot.

Operational teams should treat vendor announcements as triggers for regression tests, not as finish lines.

Document who approved production changes, which prompts were pinned, and what rollback looked like when the first anomaly appeared.

Security and compliance stakeholders need plain-language briefings: what data leaves the perimeter, what logs exist, and what happens on failure.

More in science

Comments

Loading comments…

Across the Network