OSS Supply Chain Risk Scoring - Where abandoned packages come to rest.
Ossuary is a governance-risk triage system for open source packages. It scores each package on observable governance evidence (0-100) so that a reviewer can decide where limited attention should go before an incident becomes known.
Ossuary targets the subset of supply chain attacks where governance weakness is a precondition - social engineering takeovers, abandoned packages, governance disputes. High maintainer concentration isn't inherently dangerous (pciutils has been maintained by one person for 28 years), but combined with other signals it becomes meaningful.
| Can Detect | Cannot Detect |
|---|---|
| Social engineering takeover (xz pattern) | Account compromise (stolen tokens) |
| Abandoned packages | Dependency confusion |
| Governance disputes (left-pad pattern) | Typosquatting |
| Newcomer takeover patterns | CI/CD exploits |
| Economic frustration signals | Active maintainer sabotage |
# Install from PyPI
pip install ossuary-risk
# Set GitHub token for API access (optional but recommended)
export GITHUB_TOKEN=ghp_xxxxxxxxxxxxx
# Initialize database
ossuary init
# Score a single package
ossuary score event-stream -e npm
ossuary score numpy -e pypi
ossuary score serde -e cargo
# Score with historical cutoff (T-1 analysis)
ossuary score event-stream -e npm --cutoff 2018-09-01
# Score an entire dependency tree
ossuary score-deps transformers -e pypi
# Score every component in an SBOM (CycloneDX or SPDX)
ossuary score-sbom product.cdx.json
ossuary score-sbom product.spdx.json --enrich enriched.cdx.json --annex-vii report.json
# Estimate the implied maximum CRA support period
ossuary support-period lodash -e npm
ossuary support-period-sbom product.cdx.json
# Show dependency tree with risk scores
ossuary deps express
# Generate xkcd-2347 tower visualization
ossuary xkcd-tree transformers -e pypi --tower -o tower.svg
# Batch score from seed file (--repo-aware groups sibling packages
# that share a repo so the snapshot cache is warmed once per repo)
ossuary seed-custom my-seed.yaml --repo-aware
# Show packages with biggest score changes
ossuary movers
# Show cache state
ossuary cache-statsnpm, PyPI, Cargo, RubyGems, Packagist, NuGet, Go, GitHub
Final Score = Base Risk + Activity Modifier + Protective Factors
(20-100) (-30 to +20) (-105 to +45)
Base Risk from maintainer concentration. Activity Modifier rewards active maintenance, penalizes abandonment. Protective Factors include maintainer reputation, funding (GitHub Sponsors), org ownership, visibility (downloads/stars), community size, and responsibility-shift detection.
Responsibility-shift detection: compares a contributor's recent commit share against its historical baseline. A shift from 2% to 50% on a mature project raises an alert that deserves review; it does not imply malicious intent, only that responsibility changed enough to justify a look. Guards prevent false positives for established contributors, long-tenure maintainers, and internal org handoffs.
See the methodology for full details.
The xkcd-tree command generates dependency tower diagrams inspired by xkcd 2347. Block color = risk score, block width = contributor count, arrow = most structurally critical dependency.
ossuary score-deps transformers -e pypi # score all deps first
ossuary xkcd-tree transformers -e pypi --tower -o tower.svgOssuary plugs into a Cyber Resilience Act (Regulation (EU) 2024/2847) workflow:
ossuary score-sbomconsumes the manufacturer's SBOM (CycloneDX or SPDX) and scores every component, mapping CRA Article 13(5) due-diligence on third-party components to a per-component governance signal.--enrichwrites the SBOM back with scores attached as CycloneDXcomponents[].properties[]entries or SPDX 2.3 package-levelannotations[]entries (validated against the official SPDX 2.3 JSON Schema in CI; an optional interop test round-trips throughspdx-tools).--annex-viiproduces a structured, timestamped, methodology-versioned record suitable for inclusion in the Annex VII technical documentation required by Article 13(4).ossuary support-period[-sbom]derives the implied maximum support period a manufacturer can defensibly claim under Article 13(8), bounded by the worst-governance critical dependency.
These outputs do not change the underlying scoring methodology; they are derivations on top of it. See methodology §12 for full details and the heuristic mapping behind the support-period derivation.
# Install with dashboard dependencies
pip install "ossuary-risk[dashboard]"
# Run dashboard
ossuary dashboardFeatures: risk overview, ecosystem breakdown, package detail with score history, delta detection (biggest movers).
ossuary apicurl http://localhost:8100/score/pypi/flask
curl http://localhost:8100/check/npm/expressInteractive docs at http://localhost:8100/docs.
The API binds to 127.0.0.1 by default. If you deliberately expose it with
--host 0.0.0.0, put it behind an authenticated, rate-limited reverse proxy:
scoring performs outbound registry/GitHub requests and repository clones.
Validated on 184 packages across 8 ecosystems with a formal scoped validation framework (methodology v6.4.3; current-state cutoff 2026-08-15, fixed per-incident T−1 cutoffs retained):
| Metric | All incidents | In-scope only (Scope B) |
|---|---|---|
| Accuracy | 83.7% | 92.6% |
| Precision | 94.7% | 94.1% |
| Recall | 56.2% | 76.2% |
| F1 Score | 0.706 | 0.842 |
Ossuary detects governance risk, not all supply chain attacks. The "in-scope" metrics count only incidents where governance weakness was observable before the attack — governance decay, protestware, weak-governance compromise, and governance risk. Out-of-scope incidents (credential theft on healthy projects, CI/CD exploits) are included in the dataset but not counted against recall.
2 false positives (jsonwebtoken, Newtonsoft.Json). The 10 in-scope
false negatives are listed with package-level mechanisms in the validation
report. The v6.4.3 measurement correction excludes standard GitHub [bot]
commits from the activity count; this moves telnyx from 55 to 85 because 502 of its 513
current-year commits were authored by stainless-app[bot]. Commit frustration
is also restricted to the top non-[bot] contributor; four false technical-phrase
hits disappear without changing a final score or classification.
See the validation report for full analysis. The separate post-freeze diagnostic ML experiment publishes its sanitized matrix, fixed model protocol, OOF predictions, and reviewed same-population comparisons. It is methodological support, not an independent prospective validation set.
git clone https://github.com/anicka-net/ossuary-risk.git
cd ossuary-risk
python -m venv .venv && source .venv/bin/activate
pip install -e ".[dev,dashboard]"
cp .env.example .env # add GITHUB_TOKEN
ossuary initGITHUB_TOKEN=ghp_xxxxxxxxxxxxx # GitHub API access (recommended)
DATABASE_URL=sqlite:///ossuary.db # Default; supports PostgreSQL
OSSUARY_CACHE_DAYS=7 # Score freshness thresholdMIT
MBA thesis research on OSS supply chain risk (VŠE Prague, due Dec 2026).
AI assistance declaration: AI agents from several providers were used extensively for implementation, research, validation, adversarial review and thesis drafting. Substantial repository contributions are attributed in Git history. The full workflow, verification process and author contribution are documented in the thesis AI-use appendix.
Key contribution: governance-based risk indicators are observable in public metadata before incidents occur, supporting governance-risk triage across a declared attack subset rather than a universal detector.
This repository accepts AI agent contributions. See AGENTS.md
for the repository contract and spec/ for the machine-facing version.
The contract emphasizes correctness, reproducibility, and academic honesty:
agents must not overclaim results, fabricate sources, or let methodology,
validation, code, and public documentation drift out of sync.