CryptoProver: AI System Successfully Verifies Critical Cryptographic Libraries
Key Takeaways
- ▸AI language models can automatically synthesize formal proofs for complex cryptographic libraries, reducing verification time from months to hours
- ▸CryptoProver verified critical crypto libraries (curve25519, chacha20) used by Signal and Shadowsocks with strict trust and isolation safeguards
- ▸The system achieved formal verification at $467 in API costs and 11.4 hours, versus 8 months of human effort for comparable work
Summary
Researchers have developed CryptoProver, an AI-based system that automatically synthesizes formal proofs and internal specifications for cryptographic libraries without modifying their executable code. The breakthrough demonstrates the feasibility of using language models to verify production-grade cryptography, a critical infrastructure problem that has resisted formal verification efforts.
CryptoProver has successfully generated a new independent verification of curve25519-dalek and verified RustCrypto's previously unverified chacha20 implementation against RFC 8439 specifications. These cryptographic libraries underpin globally deployed systems including Signal (218M+ downloads) and Shadowsocks. The verification process required 11.4 hours of computation at a cost of $466.99 in API fees. A parallel human-led verification of curve25519-dalek took eight months with five contributors, highlighting the efficiency gains from AI-assisted proof synthesis.
The system follows a trust-first design principle, incorporating mechanical gates that reject specification weakening, invented axioms, and cross-module failures while preventing reference proof retrieval from external sources like git history. By combining high-level API contracts with a trusted library of field specifications and arithmetic facts, CryptoProver demonstrates that AI can produce formally verified cryptographic code suitable for real-world deployment.
- Trust-first design with mechanical gates prevents specification weakening and invented axioms, maintaining cryptographic integrity
Editorial Opinion
This research suggests AI can finally unlock formal verification for production cryptographic libraries—a long-standing challenge given crypto's criticality and complexity. If CryptoProver's trust guarantees hold up to scrutiny, this could transform how we ensure the safety of cryptographic infrastructure underlying billions of devices. The efficiency gains alone (hours vs. months) are remarkable, though the reliance on external LLM APIs for critical security verification raises important questions about reproducibility and long-term maintainability.
