ContextSutra AI

07. Security and legal

This page describes engineering controls. It is not legal advice and does not claim any compliance certification.

Security boundaries

  1. No private IDE memory scraping, no chat history, no cookies, no tokens.
  2. Minimum data: only an approved root; documents on an allow-list; secret files and private directories skipped; symlinks never followed.
  3. Repository text is untrusted data. It is quoted, never promoted to an instruction.
  4. No hidden network calls, no telemetry, no model calls.
  5. No execution of repository scripts. Script names are listed only.
  6. Server is loopback-only, token-protected and read-only.

Threat analysis (STRIDE summary)

ThreatExampleControlEvidence
SpoofingAnother web page calls the local serverBearer token, Host header guardAudit V2, V4
TamperingCapsule edited after reviewCanonical SHA-256 digest; validateAudit D3, D4
RepudiationWho shared whatOut of scope for the local tool; no audit logDocumented limit
Information disclosureSecret in README ends up in a promptRedaction, file deny-list, review gateAudit P1 to P6, G2
Denial of serviceHuge or deep treeFile, depth and size limitsTests
ElevationREADME tells the AI to exfiltrateQuarantine, untrusted-data banner, skills forbid itAudit I1 to I3
Path traversal../ in a source pathisSafeRel, safeFile realpath checkAudit X1, X2
Supply chainMalicious dependencyZero dependencies, no install scriptsAudit S4, S5
Cross-site scriptingHostile heading in the HTML reportEscaping, URL scheme filterAudit S6, tests

Residual risks

Responsible disclosure

Report vulnerabilities privately through GitHub security advisories on the repository. Do not include real secrets. See SECURITY.md.

Licensing

Privacy

The tool processes only local files you point it at and writes only where you tell it. It collects nothing and sends nothing. If you share a capsule, you are responsible for what it contains and for your own data-protection duties (GDPR, contracts, internal policy).

What a reviewer should still do

Read the capsule. Run the audit. Check your own repository for personal data. Decide who may receive the file.