Skip to content

Security: MrPredic/indelible

Security

SECURITY.md

Security Policy

Supported versions

Version Supported
0.1.x

Reporting a vulnerability

Please do not open a public GitHub issue for security vulnerabilities.

Email: security report via GitHub private vulnerability reporting at
https://github.com/MrPredic/indelible/security/advisories/new

We aim to respond within 72 hours and release a patch within 14 days for confirmed issues.

Trust model

The signature provides tamper-evidence relative to a pinned public key — it is not self-authenticating.

  • indelible init writes the private key to ~/.indelible/key.pem (kept out of the repo) and the public key to indelible.pub (committed).
  • indelible.pub is the trust anchor. verify checks the .sig against it (or against --pubkey). The signature and the pinned key are deliberately kept apart: a public key that ships next to the signature is worthless, because an attacker who edits the fingerprint can re-sign with their own key and attach their own pub.
  • Rotating the pinned key is therefore a visible, reviewable event (a git diff on indelible.pub, or a re-pin in the consumer's out-of-band store).
  • Cross-party verification: a consumer verifying a third party's fingerprint MUST obtain that party's public key out-of-band (--pubkey vendor.pub) — never from the same repository that carries the fingerprint, or the anchor and the artefact share a fate.
  • What it does NOT guarantee: authenticity + integrity of origin, not the honesty of the attester. A malicious attester can sign a truthful-looking-but-misleading baseline. indelible proves who produced the fingerprint and that it was not altered after signing — not that the baseline is meaningful.

Scope

  • Ed25519 key handling (indelible/attest.py, indelible/verify.py)
  • Fingerprint signature verification bypass
  • Path traversal in IndelibleConfig.from_toml() or CLI file arguments
  • Dependency vulnerabilities in cryptography >= 42

Out of scope

  • Issues in optional extras ([deep])
  • Prompt injection in the models you are attesting (indelible is a measurement tool, not a firewall)

There aren't any published security advisories