raven-zero-day-threat-patterns

name

raven-zero-day-threat-patterns

description

Curate, query, and emit Raven's catalog of 0-day threat patterns. Use when the user says "0-day pattern", "exploit pattern library", "threat fingerprint", "variant template", "bug-class taxonomy", "match this finding against known patterns", or asks to extend the YARA / Semgrep / ARES rule packs with a new 0-day archetype. Every pattern record terminates at a tool oracle (𝒯) or a classical-ML detector (𝓜) — never raw LLM speculation — and ships with the CWE plus D3FEND Harden / Detect technique ids it counters.

Raven — 0-Day Threat Patterns

This skill is the curator of Raven's pattern library. It does not hunt and it does not fix — it owns the catalog that the hunting, detection, investigator, prevention, and fixing skills all read from. Treat patterns as durable assets: every pattern in the library MUST be reproducible, machine-checkable, and bound to a CWE plus one or more D3FEND techniques.

When to use

Trigger when the user asks any of:

• "Show me Raven's 0-day patterns for <bug class>"
• "Add a new pattern for CVE-YYYY-NNNNN"
• "Match this finding against the pattern library"
• "Promote this hypothesis into a permanent pattern"
• "Export the pattern pack for ARES / Semgrep / YARA"
• "What patterns cover CWE-<id>?"

Do not trigger for: running a hunt (use raven-zero-day-hunter), validating a single finding (use raven-zero-day-investigator), or generating mitigations (use raven-zero-day-fixing).

Inputs

Required:

1. Mode — one of: list, query, add, match, export
2. Filter (for list / query) — any of: bug_class, cwe, d3fend_id, target_type, seed_cve

For add:

• seed_cve (optional) — CVE to derive the pattern from
• bug_class — must be one of the canonical classes (memory-corruption, integer-overflow, race-condition, auth-bypass, deserialization, type-confusion, signature-malleability, account-confusion, oracle-manipulation, deserialization, ssrf, prompt-injection)
• target_type — solana-program, evm-contract, c-cpp-source, rust-source, binary-elf, binary-pe, android-apk, linux-kernel-module, web-app, llm-service
• detector — at least one tool-oracle rule (ARES / Semgrep / YARA / Nuclei template / IsolationForest feature vector) — patterns without detectors are rejected

Pattern record schema

Every entry in the library is a JSON record with these fields. Records live under raven/patterns/<bug_class>/<pattern_id>.json plus a denormalized raven/patterns/index.parquet for fast queries.

{
  "pattern_id": "P-2026-0042",
  "title": "Solana CPI privilege escalation via missing owner check",
  "bug_class": "auth-bypass",
  "cwe": "CWE-862",
  "target_type": "solana-program",
  "seed_cves": ["CVE-2024-XXXXX", "CVE-2024-YYYYY"],
  "preconditions": [
    "Program performs CPI to caller-controlled account",
    "No assert_eq!(account.owner, expected_program_id)"
  ],
  "detectors": [
    {
      "kind": "ares",
      "rule_id": "ares.solana.owner_check_missing",
      "evidence_kind": "static-AST"
    },
    {
       
       
       
    
  
     
   
   
   
   
   
     
     
     
  

Records that lack detectors or d3fend_countermeasures are invalid and MUST be rejected at write time.

Pipeline — mode-specific

Mode list and query

1. Load raven/patterns/index.parquet.
2. Apply filters (DuckDB query inline — no full-table scan).
3. Return matching pattern_id + title + bug_class + cwe + d3fend_countermeasures + true_positive_rate_observed.

Limit default 50, configurable up to 500.

Mode add

1. Validate bug_class is in the canonical list.
2. Validate every detector.rule_id exists in the corresponding tool pack (call into raven/tools/ares.py, raven/tools/yara_scanner.py, etc. to confirm the rule loads).
3. Run raven.d3fend.api.lookup_technique on every entry in d3fend_countermeasures — reject if any id is not in the ontology.
4. Run raven.d3fend.api.cwe_to_d3fend(record.cwe) — at least one returned technique MUST overlap with d3fend_countermeasures. This is the cross-check that prevents arbitrary D3FEND IDs from being attached.
5. Persist as JSON, regenerate index.parquet, commit with a structured message: pattern: add P-YYYY-NNNN <title>.

Mode match

1. Accept a finding (from raven-zero-day-investigator) as input.
2. Cross-reference its cwe, target_type, and evidence_features against the library.
3. Score each candidate pattern with cosine similarity of the evidence feature vector against the pattern's exemplar vector (raven/ml/zero_day_detector.py provides the embedding).
4. Return top-K patterns with score ≥ 0.65 and never return a match below 0.50 — below that, emit no_match rather than a noisy false positive.

Mode export

Exports a deterministic, version-tagged pack:

• pack/ares/ — ARES rules
• pack/semgrep/ — Semgrep rules
• pack/yara/ — YARA rules
• pack/nuclei/ — Nuclei templates (own-asset use only — defender edition)
• pack/manifest.json — pin SHAs, pattern ids, D3FEND coverage delta

Pack version is the git short-SHA of the patterns directory at export time.

CDP grounding — non-negotiable

Patterns are static, but the skill's emissions are not. Every match result and every add validation MUST be grounded:

• 𝒯 — the actual ARES / Semgrep / YARA rule must compile and load. The skill verifies by invoking the tool oracle on a known-good fixture and a known-bad fixture before accepting the pattern.
• 𝓜 — raven/ml/zero_day_detector.py provides the feature embedding used for similarity scoring. Embeddings produced from LLM-only output are rejected — the embedding pipeline must read the actual source artifact.
• 𝓛 — when promoting a hypothesis to a pattern, the hypothesis must carry validator evidence from at least three independent hunts (validator_runs ≥ 3). Single-shot hypotheses do not become patterns.

Defender-only enforcement

The skill MUST NOT accept patterns whose detectors reference excluded modules: raven/redteam/offensive.py, raven/tools/metasploit_integration.py, raven/tools/empire_client.py, raven/tools/exploitdb*.py. Any such reference fails validation. The skill MUST NOT export an "exploit recipe" form — the pack format is detector-only.

Output contract

For list / query / match:

# Pattern query — <filter summary>

| Pattern | Title | CWE | D3FEND | TPR | FPR | Last validated |
|---------|-------|-----|--------|-----|-----|----------------|
| P-2026-0042 | Solana CPI missing owner check | CWE-862 | D3-SCH, D3-IVV, D3-AH | 0.91 | 0.07 | 2026-05-12 |
| ... |

For add:

# Pattern added — P-YYYY-NNNN

- Title: <title>
- CWE: <cwe>
- D3FEND: <id list>
- Detectors loaded: <n>
- Cross-check D3FEND ⇄ CWE: PASS (overlap on <id>)
- Index regenerated: <row count>
- Commit: <sha>

For export:

# Pattern pack exported — v<git-short-sha>

- Patterns: <n>
- ARES rules: <n> / Semgrep: <n> / YARA: <n> / Nuclei: <n>
- D3FEND coverage delta vs previous pack: +<n> -<n>
- Manifest: pack/manifest.json
- SHA-256: <hash>

Refusal rules

1. Refuse to add a pattern without at least one working detector.
2. Refuse to add a pattern without at least one D3FEND countermeasure id.
3. Refuse to add a pattern whose D3FEND ids do not cross-check against the CWE.
4. Refuse to export an "exploit pack" — defender edition is detector-only.
5. Refuse to promote a hypothesis with validator_runs < 3.

Related skills

• raven-zero-day-hunter — produces hypotheses; this skill curates which ones become permanent patterns.
• raven-zero-day-investigator — consumes pattern library via match mode.
• raven-zero-day-detection — consumes the exported pack as its runtime ruleset.
• raven-zero-day-fixing — reads d3fend_countermeasures to drive remediation.
• D3FEND OWL integration (raven/d3fend/) — source of truth for the technique ids that every pattern must bind to (D3FEND home).