zhcharyzhang/lara-clawaudit

LARA: Verifier-Guided Rule-Level Semantic Fuzzing for Auditing Local Agent Runtimes

Artifacts for auditing the runtime layer of local LLM agents — the prompt builder, tool dispatcher, skill loader, memory writer, permission gate and network client that mediate between model output and host actions.

Released here: the historical vulnerabilities mined from four agent runtimes, the frozen rule corpora produced by verifier-guided rule-level semantic fuzzing, and every artifact needed to reproduce the reported numbers.

Contents

Path	Paper	What it is
research/	§3	Method: semantic motifs over the mined vulnerabilities, oracle feasibility, metamorphic candidates, contract evidence matrix, security-contract IR
rules/	§3	Hand-written Semgrep rules, grouped by taxonomy category
artifacts/*_historical_mining/	§3	Mined security history per runtime, with the confirmed vulnerability set and its provenance
artifacts/rounds_final/	§5	OpenClaw rule corpus, rounds 0–8. round0/S0.json is the seed set, S_FINAL.json the frozen result
artifacts/nanobot_*, artifacts/hermes_*	§5	Migration of the OpenClaw corpus to two further runtimes, plus their fuzzing rounds
artifacts/ablation_no_semantic_mutation/	§5	The w/o semantic mutation ablation: full 8-round trajectory, protocol, freeze manifest
artifacts/baseline_comparison/	§5	Stock-tool baseline (Semgrep, CodeQL) and the query-backend composition analysis
artifacts/v1_qlcoder_metric_rescore/	§5	Per-(rule, vulnerability) metric recomputation
LARA_AUTHORITATIVE_CURRENT_STATE.md	—	Consolidated state document: verified tables, claim–evidence map, limitations

Historical vulnerabilities

Each entry is a real, fixed security issue in an agent runtime, recorded with its vulnerable revision, its fix revision and the files the fix touched. Detection is scored against the patch region, so every entry carries its own ground truth.

Runtime	Language	Vulnerabilities used for evaluation
OpenClaw	JavaScript / TypeScript	275
Nanobot	Python	22
Hermes	Python	12
OpenHands	Python	12

Rules and queries

A rule pairs a security property with the scope, entities, effect and relation that define a violation. A query is one executable CodeQL encoding of that rule; a rule may have several. Counts in the tables below distinguish the two.

Corpus	Rules	Queries	Backend
OpenClaw seed (S0)	166	166	CodeQL
OpenClaw final (R8)	180	218	CodeQL
Nanobot final	136	136	CodeQL
Hermes final	172	172	CodeQL

All four corpora are 100% CodeQL. The compiler attempts a Semgrep encoding first and falls back to CodeQL; every accepted rule requires a relation between two entities, which the Semgrep pattern language cannot express. See artifacts/baseline_comparison/REALIZATION_COMPOSITION.md.

Headline results (§5; OpenClaw, 275 historical vulnerabilities)

System	Patch Recall	Success
Stock Semgrep (273 rules)	3.3%	1.8%
Stock CodeQL (71 of 88 default-suite queries)	2.9%	0.4%
Stock union	6.2%	2.2%
Seed rules, no fuzzing	61.1%	61.1%
Query rewriting only (ablation)	62.2%	62.2%
Full system	81.5%	80.0%

The stock-CodeQL row ran 71 of 88 queries; 17 do not complete inside a full-suite invocation on this codebase and were excluded, so that row is a lower bound. The excluded list and the reason are in artifacts/baseline_comparison/.

What is deliberately not included

• A deployment study on current revisions of four runtimes, and its proof-of-concept scripts. Those candidate defects have not been disclosed to the maintainers, so they are withheld to avoid uncoordinated disclosure. The corresponding sections and table of the state document are removed.
• Execution caches (CodeQL databases, per-revision SARIF, worker state) — several hundred GB of intermediate data the reported numbers were computed from, reproducible from the frozen corpora.
• Contributor email addresses appearing in mined commit metadata have been redacted.

Reproducing

Patch Recall is vulnerability-level: an entry counts if any rule reports a site inside the mapped patch region of the vulnerable revision. Success additionally requires zero in-region reports on the fixed revision. CodeQL 2.23.5, Semgrep 1.177.0; patch regions padded 3 lines either side of each hunk.

Citation

Paper under submission; citation details will be added here on release.