Skip to content

Security policy

Supported versions

Version Supported Notes
2.x Yes The main branch and the container images built from it (ghcr.io, tags latest and sha-<commit>).
1.x No The single-origin service tagged v1.0.0. Upgrade to 2.x.
Before 1.0.0 No The original training scripts.

Fixes are made on main and shipped by the Release workflow. There are no backports to older versions.

Reporting a vulnerability

Please report vulnerabilities privately through GitHub's private vulnerability reporting, which is enabled for this repository:

  1. Open the repository's Security tab, or go directly to the new advisory form.
  2. Choose Report a vulnerability.
  3. Fill in the advisory form.

Do not open a public issue, pull request or discussion that describes a vulnerability. If you cannot use GitHub's form, open a public issue titled "Security contact request" that contains no technical details. The maintainer will then set up a private channel.

A useful report includes:

  • the affected component (for example the API service, the edge Worker, the web UI, the container image, or a workflow) and the version or commit;
  • steps to reproduce, or a proof of concept;
  • the impact you expect (what an attacker gains);
  • any fix or mitigation you suggest.

Never include real API keys, tokens, or other credentials in a report, and redact them from any logs you attach.

Testing guidelines

  • Test against a local deployment (fastapi-assistant serve, or docker compose up --build) wherever possible.
  • Do not run load tests, denial-of-service attempts, or automated scanners against the public deployment. It runs on capped free-tier capacity, and doing so harms other users.
  • Do not access, modify, or delete data that is not yours, and stop as soon as you have shown that a problem exists.

Response targets

This is a single-maintainer project. The following are goals, not guarantees:

Step Target
Acknowledge the report within 5 business days
Initial assessment and severity within 10 business days
Fix or mitigation, critical or high severity within 30 days of confirmation
Fix or mitigation, medium or low severity within 90 days of confirmation

Fixes are disclosed in a GitHub security advisory once a fixed version is deployed. Reporters are credited in the advisory unless they ask not to be.

Scope

In scope:

  • the Python service in src/fastapi_assistant/: HTTP API, authentication and API keys, admission control, request limits, storage, retrieval, and model loading (including the revision and SHA-256 pinning in model-release.json);
  • the Cloudflare Worker in deploy/cloudflare-worker/ (key checks, rate limits, answer cache, origin failover, origin lock);
  • the web UI in web/;
  • the container image built from Dockerfile, and the images published from this repository;
  • the GitHub Actions workflows in .github/workflows/ and the deployment templates in deploy/ and scripts/;
  • the dependency locks (requirements/serve.txt, package-lock.json files).

Out of scope:

  • incorrect, low-quality, or made-up model answers. The model only produces text and has no tools or data access, so a wrong answer is a quality bug: please open a bug report. See MODEL_CARD.md for known limitations;
  • prompt injection that only changes the text of the model's own answer, without crossing a security boundary;
  • fastapi-assistant eval --execute-code, which runs model-generated code by design and is documented for disposable environments only (report it if that code can run without the flag);
  • vulnerabilities in third-party platforms (Cloudflare, Google Cloud, Hugging Face, Neon, Upstash, Kaggle, GitHub). Report those to the vendor;
  • vulnerabilities in upstream dependencies with no demonstrated impact on this project. Report those upstream; tell us if you can show an impact here;
  • volumetric denial of service, social engineering, and physical attacks;
  • scanner output without a demonstrated impact.

Accepted risks

These are known, reviewed, and accepted. Reports that only restate them will be closed with a pointer to this section.

PYSEC-2026-2447: diskcache 5.6.3

  • What: unsafe deserialisation of cached values in diskcache 5.6.3. No fixed release exists.
  • Why it is present: llama-cpp-python depends on diskcache for its optional LlamaDiskCache, so the serving lock (requirements/serve.txt) includes it.
  • Why it is accepted: the service never constructs a LlamaDiskCache (or any llama.cpp prompt cache), so the vulnerable code path is not reachable.
  • How it is tracked: the pip-audit job in .github/workflows/security.yml ignores this one advisory ID (--ignore-vuln PYSEC-2026-2447). Every other advisory fails the job, and the job also runs weekly. The exception will be removed when a fixed diskcache release exists, or if the service ever starts using a disk cache.

Other accepted findings

The initial audit, docs/audit/INITIAL_AUDIT.md, records further accepted findings with their reasons, including:

  • F-15: two CodeQL alerts that are intended behaviour: keys create prints a new key once to the operator's terminal, and API keys (256-bit random values) are hashed with SHA-256;
  • F-16: /health and /ready are public;
  • F-18: CORS allows any origin, because authentication never uses cookies.

Security controls and further reading

  • docs/security/SECURITY_AUDIT.md: security audit of the hardened service.
  • docs/audit/INITIAL_AUDIT.md: baseline audit, with evidence for every finding.
  • Automated checks on every pull request: CodeQL, gitleaks (full history), Trivy on the container image (fails on fixable HIGH and CRITICAL CVEs), pip-audit on the serving lock, npm audit for the Worker and the web UI, and GitHub dependency review.