Definition
A poisoned build manifest is a build, dependency or CI file that carries text written for the AI agent that reads it, in a field the build never acts on.
Tells the agent to skip a review, trust a package, hide what it finds, or send secrets somewhere. This report names the behaviors, not the wording.
Files the package manager, the build or the CI system reads: package.json, Cargo.toml, pyproject.toml, workflow and pipeline YAML, BUILD files, lockfiles, reports and attestations.
Rules files such as AGENTS.md and CLAUDE.md are the previous report, Poisoned Rules Files. Discovery-file poisoning sits in another category. Neither was measured here.
Which files the 28 rules name
| Rule | Names this kind of file | Severity | Channels declared |
|---|---|---|---|
| GLS-BMP-001 | npm package.json | high | file, web_content |
| GLS-BMP-002 | Bazel, Buck and Pants build files | high | file, web_content |
| GLS-BMP-003 | Cargo.toml | high | file, web_content |
| GLS-BMP-004 | CMake files | high | file, web_content |
| GLS-BMP-005 | Gradle and Maven build files | high | file, web_content |
| GLS-BMP-006 | Lockfiles | high | file, web_content |
| GLS-BMP-007 | Makefile and task-runner files | high | file, web_content |
| GLS-BMP-008 | Package-manager config | high | file, web_content |
| GLS-BMP-009 | pyproject.toml and Python package metadata | high | file, web_content |
| GLS-BMP-010 | Frontend build-tool config | high | file, web_content |
| GLS-CICD-001 | CodeQL config | high | file, web_content |
| GLS-CICD-002 | Dependabot config | high | file, web_content |
| GLS-CICD-003 | Renovate config and dependency-bot PR notes | high | file, web_content |
| GLS-CICD-004 | Ansible automation files | high | file, web_content |
| GLS-CICD-005 | GitLab CI pipeline files | high | file, web_content |
| GLS-CICD-006 | GitOps controller files (Argo CD, Flux) | high | file, web_content |
| GLS-CICD-007 | Jenkins pipeline files | high | file, web_content |
| GLS-CICD-008 | Observability config | high | file, web_content |
| GLS-CMP-009 | GitHub Actions step summaries | high | file, api_response |
| GLS-CMP-010 | Renovate config | high | file, api_response |
| GLS-SCAP-001 | SARIF results files | high | file, web_content |
| GLS-SCAP-002 | SLSA and in-toto attestations | high | file, web_content |
| GLS-SCAP-003 | Status and coverage badges | medium | file, web_content |
| GLS-SCAP-004 | TUF update metadata | high | file, web_content |
| GLS-SCAP-005 | SLSA and in-toto provenance | high | file, web_content |
| GLS-SCAP-006 | Coverage dashboard HTML | critical | file, api_response |
| GLS-SCAP-007 | LCOV coverage reports | critical | file, api_response |
| GLS-SCAP-009 | Kanban task-card HTML | critical | file, api_response |
The 28 rules sit in three categories: 10 build-metadata (BMP), 10 CI and CD (CICD, CMP) and 8 attestation and report (SCAP). 24 are high severity, 3 critical, 1 medium. All 28 declare file, 23 web_content, 5 api_response; none declares message. One package.json probe fired 5 of the 28 rules, so read the table as what each rule mentions, not a classifier by file type.
What we measured
| Question | Result | Denominator |
|---|---|---|
| A hostile probe fires its own rule, file channel | 28 (27 on the first try) | 28 rules, one probe each. The first miss was our probe, not the rule: it carried no named marker. We fixed the probe, not the rule. |
| ...on the web_content channel | 23 | 23 rules that declare it |
| ...on the api_response channel | 5 | 5 rules that declare it |
| A full attack in a real-named file fires the rule named for that file type | 3 of 3 for each of package.json, Cargo.toml, pyproject.toml and BUILD.bazel | 3 attacks written by us, 4 file types, 12 files |
| ...a workflow YAML file | 0 rules named for it. 3 of 3 full attacks still fired two rules aimed at other file kinds. | 3 attacks. Coverage by shared vocabulary, not a workflow rule. |
| A full attack with one kind of wording taken out fires a class rule | 0 of 15 | 15 files. Other rules still blocked 9 of the 15, so those blocks are not these rules. |
| Real manifests flagged by a class rule | 0 | 113 real files, none edited |
| Reworded attacks (same intent, new words) allowed by the whole engine | 5 | 5 files, one per manifest type |
Every hostile probe was written by us after reading that rule, so 28 of 28 is an upper bound on what the rules can do, not how often they catch real poisoned manifests. One probe per rule. This report prints no percentage. 27 of the 28 rules need several kinds of wording to appear together; GLS-CMP-009 is built differently. The 5 reworded files fired no rule of any kind.
Where a rule did not fire
| Item | What happened |
|---|---|
| Reworded attacks, one per file kind: a workflow step name, pyproject.toml, a BUILD file comment, Cargo.toml and package.json | No rule of any kind fired on any of the 5. All 5 were allowed. n=5, written by us: a list of misses, not a miss rate. |
| A GitHub Actions workflow in general | No rule is named for it. Only the step-summary rule, GLS-CMP-009, reads a part of one. |
| A hostile sentence with one kind of wording taken out | 0 of 15 fired a class rule. 9 of the 15 were blocked by other rules; 6 were allowed. |
| A bare YAML key as the only file marker | GLS-CICD-008 did not fire on our first probe. It fired once the probe named the tool. |
| The message channel | No rule in the class declares it. |
| Manifest types beyond five | Probed once each, with the rule's own wording. Not probed with reworded or paraphrased attacks. |
False positives on real files
| File kind | Real files | Flagged by a class rule |
|---|---|---|
| package.json | 55 | 0 |
| Cargo.toml | 19 | 0 |
| pyproject.toml | 14 | 0 |
| Workflow YAML | 9 | 0 |
| Python package METADATA | 2 | 0 |
| BUILD.bazel | 1 (one file: a thin sample) | 0 |
| Other (setup.py, MANIFEST.in, Dockerfile, issue and pipeline YAML) | 13 | 0 |
| All | 113 | 0 |
| Question | Result |
|---|---|
| Same 113 files through a directory scan | Agreed: no finding in any of the three categories |
| Where the files came from | 21 from our own repositories, 2 from the Action's clean fixtures, 2 installed package METADATA files, 54 third-party package.json files from one node_modules folder, 34 other real manifests on one Mac |
| Our own Cargo.toml, pyproject.toml, BUILD files | None exist, so those three kinds are third-party clones only |
| A package.json whose description quoted an attack as an example | Fired 5 class rules (n=1). A write-up shaped like a manifest is a false-positive path. |
| Outside the class | Flagged or quarantined by rules outside the class: 12 of the same 113, listed below. Not these rules, but each is a false positive or a known flag on a real file. |
| Files | Kind | Rule that fired | Our reading |
|---|---|---|---|
| 5 | Third-party package.json: wrangler, miniflare (two copies), kv-asset-handler, sourcemap-codec | GLS-PT-001, blocked | False positive. Known gap, below. |
| 2 | Ours: action.yml and acceptance.yml in our GitHub Action | GLS-CI-007, quarantined | True false positive on our own action.yml. |
| 3 | Ours: workflow and issue-template files in our scanner repository | GLS-TS-251 (1 file), GLS-GHSA-PI-202 (2 files), blocked | False positive on plain comment and description text: 3 of 113 here. |
| 1 | Ours: the scanner Dockerfile | GLS-APIP-005, blocked | False positive on its header comments: 1 of 113 here. |
| 1 | Ours: the METADATA of sunglasses itself | 24 rules, blocked | Correct flag: the README inside it lists attack examples. |
| Known gap | Status |
|---|---|
| GLS-PT-001 and GLS-CI-007 block or quarantine real manifests | GLS-PT-001 (path traversal, file channel) blocks real package.json files whose script or tool-config fields hold relative parent paths: 213 of 1,907 distinct npm packages on one Mac (581 of 5,691 top-level files), measured Nov 9 2026 on 0.6.20. scan --dir skips node_modules, so this reaches a user who scans a single manifest or a vendored tree. GLS-CI-007 quarantines our own action.yml the same way. Present since v0.2.4; not a regression. The fix is a rule change that needs review and a false-positive run, and no release is named for it yet. Listed on the known version gaps page. |
| Three other rules block real files | GLS-TS-251, GLS-GHSA-PI-202 and GLS-APIP-005 block real workflow, issue-template and Dockerfile files on comment, URL or glob text: 20 of 298 distinct files found on one Mac (copies of our own repos and installed npm packages included), measured Nov 9 2026 on 0.6.20; 4 of the 113 files in this report. |
Not a rate for the wild. The files are not a random sample, and "clean" means real files from known tools, not read line by line. The corpus used renamed copies, so we did not test whether a real path such as .github/workflows changes anything.
What the scanner does today
| Boundary | Behavior | Status |
|---|---|---|
| Engine scan, file channel | Every probe, control and corpus file was scanned with the engine on the file channel; critical or high blocks | Run |
| Directory scan | The 113 real files through a directory scan returned the same class result as the engine scan | Run: 113 files |
| Hook, proxy and MCP front doors | Not tested on a manifest | NOT RUN |
| GitHub Action | Not tested on a poisoned manifest | NOT RUN |
What we do not catch
- Real poisoned manifests. We tested none. Real-world recall for this class is unknown, and every probe was written by us after reading the rule it was aimed at.
- Reworded attacks. 0 of 5 were caught by any rule, in this class or outside it.
- GitHub Actions workflows in general. No rule is named for them; only a step summary has one.
- Any manifest sent on the message channel. No rule in the class declares it.
- Which rules do the stopping. Other rules blocked 9 of the 15 files where a class rule stayed quiet, and we did not test the decision with this class removed.
- Manifest types we probed once: Bazel, CMake, Gradle and Maven, lockfiles, task runners, package-manager config, CI files for GitLab, Jenkins, Argo and Ansible, and the report and attestation formats.
- Files over 1 MB, archives, the hook, the proxy and the Action. None was tested for this class.
Reproduce it
pip install sunglasses==0.6.20
sunglasses scan --dir . --json on a copy of your repository.
decision allow, threat_found false, exit code 0.
raw_findings holds every rule that fired. The top-level list is a merged view, and a rule absorbed into a stronger one appears only under also_matched.
No public fixture with a poisoned manifest yet. This report prints no hostile text.
Version and method
sunglasses 0.6.20 from PyPI, installed in a clean virtual environment, run isolated with an empty home folder.
November 9, 2026.
Raw findings from the engine, not the merged display view.
Synthetic probes, and real files from one Mac. This says nothing about other phrasings or about real incidents. Numbers move when the scanner moves: pin the version.