Signed Manifests Protect Requirements in LLM Agent Handoffs
Original title: Signed but Not Safe: Requirement Integrity in Multi-Agent LLM Handoffs
Signed HMAC-SHA256 manifests reduce unsafe commits in multi-agent handoffs from 69.4% to 9.0% across 1440 trajectories.
Enterprise AI agents split requests and merge results, risking requirement loss. The study compares handoff methods using Qwen3-8B and other models. Signed manifests with deterministic verification cut violations significantly while maintaining conformant commits.
- Free-text structured JSON and signed manifest handoffs compared
- Signing defeats in-transit tampering
- Determistic verifier caught 78.7% of violations
- Extraction is the trust boundary
- One adversarial line raised violations to 37.5% in some models
- AI models
- LLMs
From the original article
1 min read at arxiv.orgEnterprise AI agents increasingly split a user's request across subagents and merge their results, and each handoff can summarize away, alter or ignore a requirement.
By Xiaobo Chen · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.