llm-security/.llm-security-ignore
Kjell Tore Guttormsen d3d0cba917
docs(readme): self-scan counts after the README's own new text (383 / 441)
The AV section added two suppressed findings to README.md itself; the
remaining 58 (WARNING 61/100) did not move. Re-measured on a fresh clone.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
2026-09-22 14:24:41 +02:00

90 lines
3.2 KiB
Text

# .llm-security-ignore — Suppress expected findings when scanning this plugin
#
# Why 383 suppressed findings? A security plugin that documents attack patterns,
# ships a malicious demo fixture, and tests against deliberately evil code will
# trigger its own scanners. This is the "scanning the scanner" paradox:
#
# - examples/ contains an intentionally malicious plugin (the demo)
# - knowledge/ documents real attack regex patterns and example URLs
# - tests/ contain deliberate taint flows and suspicious URLs as test input
# - hooks/ and scanners/ contain high-entropy regex for secret detection
#
# Every suppression below is explained. Run without this file to see all of them.
# Measured 2026-09-22 (v8.1.0) on a fresh clone: 441 findings without this file,
# 58 with it (383 suppressed).
#
# Format: SCANNER:glob or just glob (applies to all scanners)
# Scanners: UNI, ENT, PRM, DEP, TNT, GIT, NET, MEM, SCR, WFL, TRG, SIG, AST, TFA
# Demo fixture: intentionally malicious (the whole point of the demo)
examples/**
# Test files contain deliberate malicious patterns as test input
TNT:tests/**
NET:tests/**
# Knowledge base documents attack patterns with example URLs and regex
ENT:knowledge/**
NET:knowledge/**
# Hook scripts contain high-entropy regex patterns and log strings
ENT:hooks/**
# Scanner code contains regex patterns that trigger entropy detection
ENT:scanners/**
# Injection patterns module contains injection keywords (by design)
TNT:scanners/lib/injection-patterns.mjs
# Command files contain long prompt strings
ENT:commands/**
# Permission findings: clean needs write tools (by design), deep-scan uses Bash
PRM:commands/**
PRM:agents/**
# Git findings: subtree split artifacts and commit message heuristics
GIT:**
# Network: README references to OWASP, Anthropic, research papers
NET:README.md
# Network: agent docs reference example domains for documentation
NET:agents/**
# Network: supply-chain hook legitimately contacts osv.dev and socket.dev
NET:hooks/**
# Orchestrator legitimately writes log file from argv path
TNT:scanners/scan-orchestrator.mjs
# Toxic flow: plugin commands/agents have Read+Bash access by design (it's a security scanner)
TFA:commands/**
TFA:agents/**
# Network: CLAUDE.md references public repo URL
NET:CLAUDE.md
# Baseline files: generated JSON with scan results (high entropy expected)
reports/baselines/**
# Unicode: the vendored commons conformance corpus carries homoglyph and zero-width inputs on purpose
UNI:scanners/commons/conformance/**
# Unicode: Cyrillic confusables are detection patterns here (the lexicon matches the homoglyph form)
UNI:scanners/lib/injection-patterns.mjs
# Unicode: homoglyph test inputs for the string-utils folding tests
UNI:tests/lib/string-utils-homoglyph.test.mjs
# Unicode: the changelog quotes the homoglyph form a fix detects
UNI:CHANGELOG.md
# Memory: the grade-F posture fixture is poisoned on purpose (that is what grade F tests)
MEM:tests/fixtures/posture-scan/grade-f-project/**
# Signature: the supply-chain blocklist names miner packages so the install hook can block them
SIG:scanners/lib/supply-chain-data.mjs
# Not suppressed on purpose: MEM findings on CLAUDE.md (it documents the hooks, but it is the
# file a poisoning would target), and AST/TRG, which have no by-design findings here.