Skip to content

Threat Model and Limitations

TrustSight reads hostile input. The threat model starts from that fact: the recipe, the diff, the committed files and the declared metadata are all attacker-controlled until proven otherwise, and the tool must stay safe while reading them.

The adversary

The model assumes an attacker who can:

  • change a PKGBUILD and any file committed beside it;
  • declare any URL, checksum, key, maintainer identity or upstream story;
  • write text designed to make a reviewer or a parser misread the recipe; and
  • wait, re-adopt an orphaned package, or coordinate several packages.

The model does not assume the attacker can change TrustSight's own code, the signed baselines, or the machine it runs on. The Security Model states the invariants that keep that boundary, and Security Boundaries enumerates the A1-A15 invariants that keep the reading machine safe.

The trust boundary

The analysis process treats every parsed byte as untrusted:

  • the PKGBUILD is never executed;
  • a declared URL is never fetched during analysis;
  • committed manifests are parsed in a sandbox rather than the analysis process;
  • writes stay inside the data directories; and
  • a bound that drops input records a coverage gap, so an incomplete analysis cannot present an unflagged verdict.

What a result claims

A result is an assertion about evidence found in the analysed material, not a statement that a package is safe. The clauses are stated in full in What a Result Claims: the score is a sum of matched evidence, an unflagged verdict is never issued for an incomplete analysis, FATAL cannot be switched off, suppression is visible, and the prose reports what changed rather than only what fired.

Limitations

Some limits are inherent to auditing a recipe without building it:

  • What the build will actually fetch can be computed at build time; the static text may not contain the URL.
  • What a binary does cannot be learned from its name or its declared intent. The optional artifact lane inspects a built manifest, outside the static score.
  • What an upstream will serve tomorrow is not visible in today's diff.
  • What a reviewer will notice is not modeled. TrustSight reduces the surface a person must read; it does not replace the person.

The reasoned ceiling, including which limits are inherent and which are merely unfixed, is in Limitations and Uncertainty. Coverage gaps and their effect on the verdict are in the Security Model.