Security Threat Model Skill

Overview

This skill executes a structured, phase-gated security threat model workflow that scans the toolkit installation for attack surface exposure, supply-chain injection patterns, and learning DB contamination. It follows the toolkit's four-layer architecture: deterministic Python scripts perform all checks and produce JSON artifacts; Phase 5 (synthesis only) is the LLM step. Each phase gates on artifact validation before proceeding.

Outputs are saved to security/ with a shared run_id for correlation across phases. Phase 5 produces an actionable threat model document.

---

Instructions

Phase 1: SURFACE SCAN

Goal: Enumerate the active attack surface of the current installation.

Create the security/ output directory and run the surface scan script:

mkdir -p security
python3 scripts/scan-threat-surface.py --output security/surface-report.json

This script enumerates:

• Registered hooks (from ~/.claude/settings.json) with file paths and event types
• Installed MCP servers (from ~/.claude/mcp.json and .mcp.json)
• Installed skills (from skills/) with allowed-tools entries
• Any file in hooks/, skills/, or agents/ containing ANTHROPIC_BASE_URL

Validate output:

python3 -c "import json; d=json.load(open('security/surface-report.json')); print('hooks:', len(d.get('hooks',[])), '| skills:', len(d.get('skills',[])), '| mcp_servers:', len(d.get('mcp_servers',[])))"

Gate (ARTIFACT VALIDATION): security/surface-report.json must exist, parse as valid JSON, and contain hooks, skills, and mcp_servers keys. A missing directory is handled gracefully with empty arrays. All artifacts are written to security/ before gating. Do not proceed to Phase 2 until this gate passes.

---

Phase 2: DENY-LIST GENERATION

Goal: Produce a concrete deny-list config derived from Phase 1 findings.

Generate the deny-list from the surface report:

python3 scripts/generate-deny-list.py \
    --surface security/surface-report.json \
    --output security/deny-list.json

The script applies these mappings from surface findings to deny rules:

• Hook uses curl or wget → append "Bash(curl *)" and "Bash(wget *)"
• Hook uses ssh or scp → append "Bash(ssh *)" and "Bash(scp *)"
• Skill allowed-tools contains unscoped Read(*) or Write(*) → add path-scoped deny entries
• Any file contains ANTHROPIC_BASE_URL override → append "Bash(* ANTHROPIC_BASE_URL=*)"

Always includes static baseline deny rules for credentials and privileged operations:

["Read(~/.ssh/**)", "Read(~/.aws/**)", "Read(**/.env*)",
 "Write(~/.ssh/**)", "Write(~/.aws/**)",
 "Bash(curl * | bash)", "Bash(ssh *)", "Bash(scp *)", "Bash(nc *)",
 "Bash(* ANTHROPIC_BASE_URL=*)"]

Display deny-list for human review:

python3 -c "
import json
d = json.load(open('security/deny-list.json'))
print('Deny-list entries to add to settings.json:')
for rule in d['permissions']['deny']:
    print(' ', rule)
print()
print('Review security/deny-list.json before merging.')
"

Gate (HUMAN APPROVAL REQUIRED): The deny-list is produced for human review only — it is never merged automatically. Display the diff and block until the operator confirms review. This gate is the highest-ROI control in the workflow. In --ci-mode, skip this gate and proceed to Phase 3. Do not proceed without explicit acknowledgment.

---

Phase 3: SUPPLY-CHAIN AUDIT

Goal: Scan all installed hooks, skills, and agents for injection patterns and hidden characters.

Run the supply-chain audit:

python3 scripts/scan-supply-chain.py \
    --scan-dirs hooks/ skills/ agents/ \
    --output security/supply-chain-findings.json

Detection patterns (full regex details in scripts/scan-supply-chain.py source):

Pattern	Severity
Zero-width + bidi Unicode characters	CRITICAL
HTML comments and hidden payload blocks	CRITICAL
ANTHROPIC_BASE_URL override in any file	CRITICAL
Instruction-override and role-hijacking phrases	CRITICAL
Outbound network commands in hooks/skills	WARNING
enableAllProjectMcpServers setting	WARNING
Broad permission grants without path scoping	WARNING

Check for CRITICAL findings:

python3 -c "
import json, sys
d = json.load(open('security/supply-chain-findings.json'))
crits = [f for f in d.get('findings', []) if f.get('severity') == 'CRITICAL']
warns = [f for f in d.get('findings', []) if f.get('severity') == 'WARNING']
print(f'CRITICAL: {len(crits)}, WARNING: {len(warns)}')
if crits:
    for c in crits:
        print(f'  CRITICAL: {c[\"file\"]}:{c.get(\"line\",\"?\")} -- {c[\"pattern\"]}')
    sys.exit(1)
"

Gate (BLOCKING CRITICAL FINDINGS): Any CRITICAL finding halts forward progress. All CRITICAL findings must be remediated or explicitly acknowledged before Phase 4 can start. This includes zero-width Unicode, ANTHROPIC_BASE_URL overrides, hidden payloads, and instruction-override phrases. WARNING findings are logged but do not block. Log warnings in the threat model under "Gaps and Recommended Next Controls" with acceptance rationale.

---

Phase 4: LEARNING DB SANITIZATION

Goal: Inspect the learning DB for entries that may contain injected content from external sources.

Run the sanitization check in dry-run mode (never mutates without explicit --purge):

python3 scripts/sanitize-learning-db.py \
    --output security/learning-db-report.json

Flags entries where:

• key or value contain instruction-override or role-hijacking phrases
• source is pr_review, url, or external (high-risk origins)
• value contains zero-width Unicode or base64 blobs
• first_seen is older than 90 days and source indicates external origin

Review flagged entries:

python3 -c "
import json
d = json.load(open('security/learning-db-report.json'))
flagged = d.get('flagged_entries', [])
print(f'Total flagged: {len(flagged)}')
for e in flagged[:10]:
    print(f'  [{e[\"severity\"]}] id={e[\"id\"]} source={e.get(\"source\",\"?\")} action={e[\"action\"]}')
if len(flagged) > 10:
    print(f'  ... and {len(flagged)-10} more. See security/learning-db-report.json')
"

Gate (DRY-RUN BY DEFAULT): The script operates in dry-run mode by default. No rows are deleted without explicit operator request and --purge flag. Present the report to the operator. If purge is desired after review, re-run with --purge. Learning DB is not found gracefully — script produces an empty report (total_entries: 0, flagged_entries: []). Proceed to Phase 5 when operator acknowledges the report or when no entries are flagged.

---

Phase 5: THREAT MODEL SYNTHESIS

Goal: Synthesize Phases 1-4 findings into an actionable threat model document. This is the only LLM-driven phase.

Load all phase artifacts:

• security/surface-report.json
• security/deny-list.json
• security/supply-chain-findings.json
• security/learning-db-report.json

Write security/threat-model.md with these required sections (validator checks for exact headings):

# Threat Model

## Run Metadata

## Attack Surface Inventory

## Active Threats

## Mitigations In Place

## Gaps and Recommended Next Controls

## Deny-List Status

## Supply-Chain Audit Summary

## Learning DB Sanitization Summary

Write security/audit-badge.json:

{
  "status": "pass",
  "timestamp": "2026-01-01T00:00:00Z",
  "run_id": "from-surface-report",
  "critical_count": 0,
  "warning_count": 0,
  "phases_completed": 5
}

Status is fail if any CRITICAL finding was not remediated or if any phase gate did not pass.

Validate outputs:

python3 scripts/validate-threat-model.py \
    --threat-model security/threat-model.md \
    --badge security/audit-badge.json

Gate (ARTIFACT VALIDATION WITH RETRY LIMIT): validate-threat-model.py must exit 0. If validation fails, add the missing sections and re-run. Maximum 3 fix iterations before escalating to operator for review.

---

Error Handling

Supply-chain audit CRITICAL finding blocks progress

Cause: A hook, skill, or agent contains zero-width Unicode, ANTHROPIC_BASE_URL override, or known injection phrase.

Resolution:

1. Open the flagged file at the reported line number
2. Determine if it is a legitimate false positive (e.g., documentation discussing injection patterns)
3. If false positive: add the file to --exclude list and re-run scan-supply-chain.py
4. If genuine: remediate the file (remove hidden payloads, instruction-override phrases) before continuing to Phase 4

Validation fails with missing sections

Cause: Phase 5 synthesis omitted a required section heading.

Resolution: Read the validator output for the exact missing section name. Add the section to security/threat-model.md with content synthesized from the phase artifacts and re-run validate-threat-model.py. Maximum 3 fix iterations before escalating to operator.

Missing configuration or databases

Cause: ~/.claude/settings.json or learning DB doesn't exist.

Resolution: These are handled gracefully by the scripts:

• Missing settings.json → surface-report produces empty arrays for hooks
• Missing learning DB → sanitization report returns total_entries: 0 and flagged_entries: [] These are not error conditions. Re-run with --verbose for detail on missing paths.

---

References

• ADR-102: Security Threat Model Skill
• pretool-prompt-injection-scanner.py -- session-time injection scanner (complements, does not replace this skill)
• learning_db_v2.py -- learning DB schema and connection interface
• OWASP MCP Top 10 (living document)
• Snyk ToxicSkills research: 36% of public skills contained injection patterns