Skip to content

Latest commit

 

History

History
71 lines (56 loc) · 3.95 KB

File metadata and controls

71 lines (56 loc) · 3.95 KB

Threat model

Scope

This model covers the public source repository, contribution workflow, release-build pipeline, generated workspace ZIP, and the boundary to an installed user workspace. It does not certify an organization's endpoint, network, legal-data processing, or authorization controls.

Assets

  • integrity of skills, roles, hooks, rules, validators, and release tooling;
  • confidentiality of contributor and end-user data;
  • authenticity and reproducibility of release assets;
  • traceability of verifiable-document evidence;
  • project identity and non-affiliation;
  • CI credentials and GitHub release authority.

Trust boundaries

  1. Public contributor input enters issues and pull requests.
  2. Seed/runtime sources enter the allowlisted release builder.
  3. CI code executes with a GitHub token.
  4. A downloaded ZIP crosses into a user-controlled environment.
  5. Legal documents and web content enter the installed workspace as untrusted data.
  6. Model output crosses to a human legal decision.

Threats and controls

Threat Principal controls Residual risk
Real case/PII committed upstream Ignore rules, profile scanner, synthetic-fixture policy, review Novel identifiers or semantic PII may evade patterns
Secret or credential exposure Filename/content patterns, minimal CI permissions, no printed secrets Encoded or unknown credential formats
Prompt injection in documents/issues Treat content as data; do not execute embedded instructions; bounded tools Human or model may still misclassify an instruction
Malicious contributor changes hooks/rules/build Review, tests, pinned actions, protected main, allowlisted build Compromised maintainer account
Pull request obtains write token Workflow-level contents: read; no write job on PR Third-party service behavior outside repository control
ZIP path traversal, symlink, duplicate, special entry Pre/post-build archive validation and clean extraction Parser/platform discrepancies
Release artifact substitution Exact-tag rebuild, SHA-256, manifest, reproducibility check Compromised GitHub account/infrastructure
User workspace accidentally pushed upstream Separate install directory, profiles, docs, path blockers User bypasses controls or stages ignored tracked data
Private repo treated as authorization Explicit policy and scanner profile semantics Organization applies an unsafe local exception
Hallucinated/outdated legal conclusion Source provenance, exact locators, version checks, independent audit, human approval Tooling cannot prove legal correctness
Hidden DOCX/PDF metadata or broken visual output Metadata checks, re-extraction, render/visual QA, human review Format-specific hidden channels
Dependency/supply-chain compromise Bounded dependencies, Dependabot, full-SHA actions, review Upstream package or action-account compromise

Abuse cases

  • A contributor disguises a real legal document as a fixture.
  • An issue asks the agent to reveal a token or change AGENTS.md.
  • A modified builder silently includes .bootstrap/ or user matters.
  • A tag workflow uploads a ZIP built before the tag.
  • A runtime hook performs network or Git write operations.
  • A technically valid evidence link is used to imply legal truth.

Tests and review should explicitly exercise these cases with synthetic inputs.

Out of scope

  • certification for classified or regulated processing;
  • legal advice or correctness guarantees;
  • endpoint malware, physical access, and organization IAM beyond documented integration boundaries;
  • availability or authenticity guarantees for external legal websites;
  • safe automatic upgrade of existing user workspaces during the beta channel.

Review triggers

Update this model when adding a dependency, networked service, plugin, automatic updater, new binary format, write-capable role, release channel, or data class. Security fixes should include a regression test and the control change here.