58 lines (40 loc) · 2.1 KB

Threat Model

This document defines what the WatchLLM kernel must defend against on the local write path.

Threat Categories

1) Hardcoded Secrets

Example: const apiKey = "sk_live_...";
Risk: credential exposure in source control and artifact pipelines.
Deterministic rule mapping: secret-literal rule with AST context checks.

2) Unsafe Environment Handling

Example: assigning a known secret value literal instead of reading from env.
Risk: bypass of safe secret retrieval patterns.
Deterministic rule mapping: secret rule allows known safe retrieval contexts and blocks assignment-like dangerous contexts.

3) Forbidden Imports

Example: import { execSync } from "child_process";
Risk: arbitrary command execution or sandbox escape.
Deterministic rule mapping: forbidden-import rule with explicit denylist.

4) Unsafe File System or System Access

Example: importing privileged modules where policy forbids them.
Risk: destructive file operations or unauthorized host interaction.
Deterministic rule mapping: policy-driven forbidden module and path checks.

5) Service Boundary Violations

Example: auth service importing db/internal directly.
Risk: coupling, privilege leakage, and policy bypass.
Deterministic rule mapping: boundary rule using declared allowed/forbidden import surfaces.

6) Auth Flow Violations

Example: protected DB mutation before auth guard in handler scope.
Risk: unauthorized data mutation.
Deterministic rule mapping: auth-flow call-order analysis in handler scope.

7) AI-Generated Bypass Patterns

Example: indirect structure that attempts to hide unsafe order or imports.
Risk: superficially valid code that violates safety intent.
Deterministic rule mapping: AST-structural checks and explicit inconclusive handling where proof is incomplete.

8) Heuristic False-Positive Pressure

Example: string match in comments or safe contexts triggering noisy blocks.
Risk: trust erosion and disablement pressure.
Deterministic rule mapping: structure-aware checks with constrained scope and fixture-backed regression tests.