LatticeGuard: Post-Quantum Audit Trails for Detecting and Reproducing Collusion in Multi-Agent AI Systems

Multi-agent AI systems increasingly use natural-language messages, tools, and shared workspaces, creating opportunities for agents to coordinate in ways that evade transcript-level oversight. Existing collusion audits can detect suspicious behavior, but they often depend on mutable logs, trusted central auditors, or white-box access to model activations. This project proposes LatticeGuard, a post-quantum audit layer that makes collusion investigations reproducible, tamper-evident, and privacy-preserving. The core idea is to combine collusion detectors with a lattice-based cryptographic evidence protocol: each agent event is committed into an append-only audit ledger using post-quantum signatures, Merkle commitments, and optional lattice-based zero-knowledge proofs for selective disclosure. Auditors can later verify that transcripts, tool calls, detector scores, and interventions were not modified, while sensitive prompts or user data remain hidden unless policy conditions require disclosure. We apply this protocol to three collusion families—steganographic coordination, hidden coalition behavior, and monitor-policy collusion—and define evaluation metrics for detection utility, audit reproducibility, tamper resistance, privacy leakage, and overhead. LatticeGuard is valuable because it addresses a gap between “detecting collusion” and “proving what happened” in a quantum-resilient way. The main takeaway is that future AI safety audits should treat cryptographic provenance as a first-class component of collusion detection, not as an afterthought.