Skip to content

Security: renweimeng/posttrain-timemachine

Security

docs/SECURITY.md

Local security boundary and dependency scans

The service binds loopback and rejects unexpected Host and cross-origin mutation requests. There is no remote sharing mode or arbitrary Python/SQL endpoint. Reward rules use a closed AST. Import validates finite values, IDs, schema, foreign keys, file/line sizes, nesting and YAML alias graphs. Snapshot and report paths are constrained; report ZIPs are verified in memory, never extracted. Captured output is rendered as text, with a restrictive production CSP.

On 2026-11-03 the production frontend (npm audit --omit=dev) and the complete default Python runtime dependency closure (pip-audit against requirements-runtime.lock) reported zero known advisories. This is a scan result, not a security certification. Public scan JSON is in docs/benchmarks/.

The optional .[trl] reference environment is deliberately pinned to TRL 0.24.0 / Transformers 4.57.1 / Torch 2.9.0 / Accelerate 1.11.0 so that numerical fixtures have a fixed executable source. Scanning the full test environment reported 16 advisory entries across 3 optional packages (Accelerate, Torch, Transformers; some aliases describe the same issue). Several concern loading untrusted checkpoints, remote model code or conversion utilities. See docs/benchmarks/pip-audit.json for exact identifiers and descriptions.

These libraries are not default PTM dependencies, and neither the CLI demo nor the server imports them. The supplied reference script runs offline with a locally constructed random model and tokenizer, no external checkpoint and no resume/conversion/custom-code path. Do not use that pinned reference extra to load unknown models or checkpoints. A production training environment needs its own current dependency review; reference parity does not establish security of an arbitrary trainer deployment. No Python advisories were suppressed to make the scan green.

Tests cover browser XSS text rendering, invalid Origin/Host, chunked request limits, unsafe paths and symlinks, archive traversal, hash tampering, YAML aliases, untrusted DSL types, missing capture inputs and preservation of trainer rewards when capture fails. No cloud service, telemetry or outbound runtime request is part of the product. Development tools and dependency installation have their own network behavior; Storybook telemetry is disabled in project config.

There aren't any published security advisories