Know what you inherit
See which inputs remain untrusted, which guarantees are safe to rely on, and which controls belong downstream.
Threat modeling for open source
Make the implicit security contract explicit.
Agent skills that turn source, documentation, and maintainer decisions into a version-bound model downstream users and vulnerability triagers can actually apply.
security-contract.yml
version bound
- 01Which inputs and adversaries are untrusted?
- 02Which properties does the project guarantee?
- 03Which properties are explicitly disclaimed?
- 04What must the caller or integrator enforce?
A contract, not a vulnerability report
Two readers, one shared model
01 / Integrators
02 / Maintainers
03 / Triagers
Security decisions become usable.
The model records not just scope, but the evidence and authority behind every claim.
Review concrete claims
React to a draft with proposed answers instead of starting from an empty security questionnaire.
Route findings consistently
Apply a closed disposition set with citations, provenance, and fail-safe escalation when the model is incomplete.
One model, three artifacts
Readable by people and tools.
Canonical prose
threat-model.md explains the contract, its evidence, open questions, and triage rules.
Structured sidecar
threat-model.yaml preserves the model structure and provenance for automated triage.
Portable export
threat-model.json provides a deliberately conservative, schema-backed interchange format.
The workflow
Draft first. Challenge it. Then publish.
Specialist skills divide the work while the orchestrator preserves one evidence trail.
3.1Orient
3.2Mine docs
3.3Map surface
3.4Interview
3.5Draft
3.6Backtest
3.7Sign off
Featured model
zlib
A native compression library with one clear untrusted surface and several easy-to-confuse caller obligations.
The example separates attacker-controlled compressed bytes from caller-trusted buffers and lengths, distinguishes checksum integrity from authenticity, and records output budgeting as a downstream duty.
Explore the zlib modelAttacker-controlled compressed bytes
trust boundary
zlib inflate APIs
untrusted output
Calling application and downstream sinks
Start with evidence
Model the project you actually ship.
Install the skills, point an agent at a checkout, and review a concrete draft.