GrantBench (development splits)

GrantBench is the benchmark used to evaluate GrantLock, a lockfile for who an AI agent stack runs as (API keys, OAuth grants, service accounts, MCP server secrets, CI tokens, cloud roles and their scopes) with a CI gate on semantic changes to those grants. Each case is a small repository snapshot.

• planted: a risk of one class G1–G8 planted into a least-privilege repository (snapshots repo/, and for dev-ext also base/ = the reviewed state before the change)
• clean: a least-privilege, fully declared repository (for false-positive measurement)
• diff_pair: before/ and after/ snapshots with the expected set of classes

Classes: G1 inline secret value · G2 scope added to an existing identity · G3 wildcard / admin grant · G4 identity without an owner · G5 identity shared by several components · G6 long-lived static key where federation is available · G7 new undeclared identity · G8 dormant grant.

Splits

Split	Cases	Surfaces	corpus_sha256
dev	120 (planted 80, clean 24, diff_pair 16)	MCP configs, GitHub Actions, .env	806d821b43fe1813f21d0c7dded45e2b5871b50426a012fd791dbf5627aeae76
dev_ext	68 (planted 48, clean 12, diff_pair 8)	docker-compose, Kubernetes, Terraform	0f5582ee1642557f779b93444f6c5653f15cb92359aa6bcfc1d37352c0244a8d

Each row holds the case metadata (case_id, kind, lang, planted, expected_labels, description) and snapshots: for every snapshot directory (in dev, repo/ and after/ include the committed grantlock.json generated from the reviewed state), the list of files with path, size, SHA-256 and content. All secret-like strings are invalid dummies that contain the marker FAKE. dev dummies imitate only the prefix of real formats; dev_ext dummies follow the real formats (prefix, length, alphabet).

Sealed holdout corpora are not part of this dataset (neither holdout v1 nor v2).

Results (GrantLock and baselines)

Corpus	GrantLock G1–G8 recall	Clean false positives	Diff accuracy	gitleaks G1	trufflehog G1	Baselines G2–G8
dev (frozen)	10/10 for every class	0/24	16/16	6/10	7/10	0/10 per class
dev_ext	6/6 for every class	0/12	8/8	5/6	6/6	0

The development splits were generated by the same author who wrote the detector; treat these numbers as consistency checks, not as evidence of performance on configurations written by someone else.

Gate history (sealed holdouts by independent authors, not included here)

Pre-registered bar (GATE 2): recall on G2, G3, G7 each ≥ 80% and mean ≥ 90%; clean false positives ≤ 5%.

Step	Result	Outcome
Holdout v1	G2 6/7, G3 5/7, G7 8/8 (mean 85.7%); clean false positives 0/16	Fail (recall)
Fix round 1	Four blind spots fixed	—
Holdout v2	G2 7/8, G3 7/8, G7 8/8 (mean 91.7%); G1 6/8 (one detector exception); clean false positives 5/16; baselines G1 5/8 each, G2–G8 0	Fail (false positives)
Fix round 2	The compose rule behind all five false positives narrowed; format-table / schema mismatch fixed	—
Re-verification on disclosed v2 (not a gate)	Clean false positives 0/16, G1 7/8, diff 8/8; G2, G3, G7 unchanged	not a pass

The pre-registered GATE 2 has not been passed.

License

Copyright © 2026 NagaYu. All Rights Reserved. Published for evaluation only; no license to use, copy, modify or distribute is granted (see LICENSE). Licensing or transfer inquiries: https://nagayu-portfolio.nagayu.workers.dev/en