Skip to content

Security: developerz-ai/ts23d

SECURITY.md

Security

Report privately. Don't open a public issue for a vulnerability.

Reporting

Use GitHub's private vulnerability reporting: Report a vulnerability. Include affected version, a minimal recipe or input that reproduces it, and impact.

Expect Within
Acknowledgement 3 business days
Assessment + plan 10 business days
Fix released Severity-dependent; critical issues first

Supported versions

0.x: only the latest published version gets fixes. No backports before 1.0.

Threat model

Surface Position
Recipes are code t3d build, t3d shot and t3d list import() the recipe module and run it with your user's permissions. Treat a recipe from a stranger like any script from a stranger: read it, or run it in a sandbox / container. There is no recipe sandbox (ADR 002)
Generated outputs (GLB, PNG) Written by our own zero-dependency writers. A malformed output that crashes a consumer is a bug worth reporting
importGlb (shipped) Parses untrusted GLB. The reader bounds accessor sizes and counts (empty accessors capped at 2^24 elements), the PNG/zip inflate is capped, and import scale is checked. Parser bugs are in scope
refine, t3d finish --model http (shipped) Talk HTTP to an endpoint you configure, with a timeout (timeoutMs: 600 s refine, 300 s finish). Replies are read through the bounded GLB reader. Cached outputs are content-hashed. Secrets are passed by the caller, never read from ambient env
Neural adapters (httpAdapter, comfyAdapter, openaiImageAdapter, kubeezAdapter, localGenerator, t3d neural run) Talk HTTP only to endpoints you configure in t3d.config.ts (which is code, like a recipe). Every request has a deadline (60 s default; whole job 30 min) and a reply cap (64 MB, checked on content-length and on the body); replies must be JSON. Model outputs are untrusted: PNGs and GLBs go through the bounded readers and task validation before use (T3D_NEURAL_OUTPUT_INVALID). Committed outputs under recipes/<name>/neural/ are capped at 12 MB of committed bytes per recipe (local outputs are git-ignored and outside the cap) and re-hashed on every load. Builds never touch the network: only t3d neural run does. Secrets go in headers / apiKey, never read from ambient env by the library, and never written to records, caches or errors; kubeezAdapter does not send its key to the download CDN (ADR 014, 015)
Network scripts (bun run textures, bun run data) Fetch with a 60 s timeout and a capped zip inflate; downloads are sha256-pinned. As of 2026-11
Supply chain Zero runtime dependencies (ADR 005). Releases publish from GitHub Actions via npm OIDC trusted publishing with provenance; no long-lived npm token exists (PUBLISHING.md)

Verifying a release

npm view @developerz.ai/ts23d@<version> dist.attestations _npmUser
npm audit signatures

A provenance attestation from developerz-ai/ts23d's release.yml means the tarball was built by CI from the tagged commit. A version without one was published by hand (only the one-time bootstrap should be).

There aren't any published security advisories