cs.CRNov 5, 2026

Signed but Not Safe: Requirement Integrity in Multi-Agent LLM Handoffs

Authors: Xiaobo Chen, Genliang Zhu, Ziyuan Wang, Chu Wang

Organizations: University of Massachusetts Amherst · Accentrust & Georgia Institute of Technology · Duke University · Accentrust & University of Illinois Urbana-Champaign

Abstract

Enterprise AI agents split requests across subagents, and each handoff can drop, alter or ignore a requirement. A preregistered study on synthetic, exactly scorable tasks compares free-text, structured JSON and signed requirement-manifest handoffs: the orchestrator extracts each requirement as a verbatim quote plus a checkable predicate, code signs it and deterministically verifies the merger's decision. With Qwen3-8B (1,440 trajectories, five threats), signed manifests cut unsafe commits under attack from 69.4% to 9.0% (60.4 pp, 95% CI [51.4, 69.4]) with no detectable utility loss versus structured handoffs (S2 -0.4 pp, 95% CI [-14.2, 13.3]). Signing defeats in-transit mutation: unsigned manifests violated 79.2 pp more often than signed ones, which never did; this replicated in gpt-oss-20b and gemma-4-12b. There, without an adversary, signed manifests were descriptively slightly worse than structured handoffs (11.1% vs 2.8%; 16.7% vs 8.3%; not tested) and lost utility overall (gpt-oss-20b, all threats: -28.7 pp). On identical Qwen3-8B trajectories, the deterministic verifier caught 78.7% of violating decisions versus 59.0% for an LLM checker, and an oracle manifest removed all remaining unsafe commits, so the residual risk is in extraction. A preregistered pre-signing injection confirmed this: one adversarial line raised signed-manifest unsafe commits from 14.6% to 37.5% (Qwen3-8B), 11.1% to 63.9% (gpt-oss-20b) and 16.7% to 97.2% (gemma-4-12b), and every such commit carried a manifest corrupted before signing. Signing protects transit; extraction is the trust boundary. A testable pre-signing specification (PSE), specified after the injection results and frozen before an offline evaluation, blocked all these injected commits (fixed templates) but refused 45.8%, 11.1% and 22.2% of benign requests, changing completions by -8.3 pp, -2.8 pp and 0.0 pp.

Figures & tables

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Explore similar work

Sep 5, 2026

From Review to Authorization: Key-Isolated Threshold Signing for LLM Agents

Autonomous LLM agents can turn semantic judgments into consequential actions such as payments and permission changes. If the same process makes these judgments and controls a reusable signing credential, a mistaken judgment can reach execution without further approval. To make LLM review govern execution, we present KITA, an authorization architecture that keeps the user's personal signing key and every threshold signing-key share outside all LLM processes. Under threshold-signature unforgeability and our system assumptions, an adversary controlling the proposer and fewer than t reviewer-signer domains cannot produce a valid authorization for a new action without signing contributions from t distinct domains. This execution-bound authorization integrity ensures that each accepted authorization includes a share from an uncompromised domain, bound to the canonical action and released only after authenticated reviewer approval. We implement an in-process prototype of KITA with a structured-output LLM adapter and threshold BLS. Six system tests validate quorum gating and message binding, seven controlled baseline scenarios compare authorization outcomes, and cryptographic microbenchmarks measure online signing cost and its scaling behavior.
Aug 25, 2026

Stale Does Not Mean Unsafe: Guard Precision for Tool-Using LLM Agents under Infrastructure State Races

Tool-using language-model agents increasingly mutate schedulers, data pipelines, object stores, and access-control systems. Between an agent's read and its commit, external state can change, but not every change makes the commit unsafe. We separate invalidating races, which break a declared safety predicate, from predicate-preserving and irrelevant races, and ask how precisely runtime guards distinguish them. Our deterministic simulator separates visible from authoritative state and injects five non-atomic failure mechanisms across 16 infrastructure tasks in four domains; frozen agent proposals are replayed counterfactually under every controller without an LLM judge. We evaluate three commit-time guard granularities (global epoch, read-set version, semantic commit predicate), multi-level verification, and model-side gates on three locally hosted quantized model families (Qwen3-4B, Phi-4-mini, Gemma4-8B; 3,456 trajectories on one GPU). All three guards eliminate unsafe commits, but their availability differs sharply: freshness-based guards needlessly block 92-95% of benign races, forfeiting up to 43% of safe task completions, while the complete predicate guard blocks none. That precision is contract-dependent: deleting a single declared clause converts exactly its fault family into unsafe commits (up to 7.9%). Model-side signals do not substitute: verbal confidence is miscalibrated (ECE approximately 0.37), action agreement matches a random gate, a cautionary prompt leaves the direct unsafe rate essentially unchanged, and after a freshness-guard block agents re-commit unsafely from refreshed but still-incomplete reads. Under degraded telemetry a hidden concurrent mutation remains observationally clean, bounding every selective policy. Precise runtime enforcement therefore requires semantic contracts, not freshness heuristics or model self-assessment.
Jul 11, 2026

Temporary Authority, Permanent Effects: Commit-Time Authorization for LLM Agents

An LLM agent often acts on an approval it observed earlier: it reads the approval, prepares the request, and commits later. If the approval was withdrawn, or covers a different effect, the service can apply the change and report success. Seventeen fixed advisories from three deployed agent products document failures: an approval did not cover the executed effect, with no intervening change (12), or survived a change in the facts it depended on (5). We propose commit-time authorization (CTA), requiring at least one complete, current justification for the exact effect at commit: a live grant, current premises, ready dependencies, an exact-effect binding, and an eligible path. CommitGate selects one complete justification. The sink validates it or consumes an expiring, single-use derived grant atomically with the effect. In a five-model service study whose prompt states the authorization rule, timely notice prevented nearly all unauthorized effects for GPT, Fable, and Sonnet, and fewer for Qwen and Gemma. Without a signal, all five mostly completed tasks with unauthorized effects, even when a read could reveal the change; Fable's counts use its provider's answered requests. Protected sinks enforced the declared policy in LangChain approval flows, OpenHands coding tasks, services under cloud contention, and a transaction-runtime integration. The services and runtime refused effects that lost authority after a positive check and preserved fresh alternatives. After refusal, GPT, Fable, and Sonnet on OpenHands mostly published to a separately authorized target. Machine-checked proofs show whole-candidate selection, trace safety for mediated atomic sinks, and exact single-use transfer. With honest issuers, unique candidate IDs, and a premise required by policy, every candidate CTA accepts in the model carries a live grant given at the current version; slot-by-slot checks do not ensure this.