Suppression cookbook

DefenseClaw applies suppressions in three places before emitting a finding. Each layer answers a different question:

Pre-judge strips

Strips run on the raw input before it ever reaches the judge model. Use them to remove predictable, harmless content the judge would otherwise classify as a leak.

pre_judge_strips:
  - id: STRIP-DOCS-EXAMPLE-AWS-KEY
    pattern: 'AKIAIOSFODNN7EXAMPLE'
    context: AWS docs example credential keeps tripping the PII judge
    applies_to: [pii]

When to use:

• A documented "fake" credential is included in an example or test fixture.
• Telemetry payloads from your own services are being flagged as exfiltration markers.
• A known tutorial pattern (e.g., the AWS docs example) shows up in user prompts often enough to drown the real signal.

When not to use:

• The pattern matches anything that could ever be real production data. Use a finding suppression instead, which can be conditional.
• The judge is wrong about a category — that's an LLM judges section problem, not a strip.

Finding suppressions

Drop a finding only when both the rule id pattern AND a per-finding entity pattern match. Optionally narrow further with a condition such as is_epoch (numeric, looks like a Unix timestamp) or is_platform_id (matches well-known platform-issued ID formats).

finding_suppressions:
  - id: SUPP-EPOCH-TIMESTAMP
    finding_pattern: '^SEC-DATA-NUMERIC'
    entity_pattern: '^\d{10,13}$'
    condition: is_epoch
    reason: Epoch timestamps look like card numbers under a numeric-only rule

Recipes you can copy from the recipe catalog:

• SUPP-EPOCH-TIMESTAMP: numeric "secrets" that are actually Unix timestamps.
• SUPP-PLATFORM-IDS: platform-issued IDs (Stripe customer ids, Slack team ids, etc.) that look like API keys to a permissive regex.

Tool suppressions

Suppress findings on tools whose name matches a regex. The wizard scopes this very tightly so you don't accidentally silence write/destructive variants.

tool_suppressions:
  - tool_pattern: '^(shell|bash|sh)\.execute$'
    suppress_findings: [JUDGE-INJ-DESTRUCTIVE]
    reason: Cosmetic shell commands generate noise without security risk

Pair with the rule pack:

1. Add a JUDGE-INJ-DESTRUCTIVE category in the LLM judges section that lights up on actual destructive calls (rm -rf, dd, mkfs, truncate).
2. Add the tool suppression here so cosmetic invocations of the same shell tool don't trip the same finding id when the model just wants to git status.

Workflow

In the Policy creator:

1. Open the LLM judges section and confirm your judge categories actually map to the findings you care about. Add or remove categories before tuning suppressions.
2. Open the Suppressions section. Use the recipe picker on each tab to seed your suppressions with the bundled strict-pack entries.
3. Use the regex tester in each suppression to verify your pattern matches the entity you intend without matching anything else. The wizard will turn the example pill green when the pattern matches.
4. Watch the findings tray (bottom of the wizard) for SUPP_OVER_BROAD warnings. Tighten any pattern the wizard flags before exporting.