agentsclimarketplace

Threat modeling

Skill ulises-jeremias/agent-toolkit/skills/agentic-security/threat-modeling

STRIDE + agentic threat modeling — architecture discovery → assets/trust boundaries/data flows/actors → STRIDE + agentic threats → risk-ranked mitigations → incremental review → security acceptance criteria. Swarm-friendly.From its SKILL.md

Install
npx -y skills add ulises-jeremias/agent-toolkit --skill threat-modeling

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

  • 19 stars19 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

SKILL.md

8.3 KB, as published. Nobody here has run it

Threat Modeling — STRIDE + Agentic Threats

Guide architecture → threat model → security review → security acceptance criteria. Produces a STRIDE table augmented with agentic-specific threats (prompt injection, tool poisoning, excessive agency, inter-agent trust), risk-ranked with mitigations and acceptance criteria.

Evidence-linked, portable, incremental. Architecture discovery reuses C4/Mermaid inputs (#369); outputs feed agentic-security-reviewer + owasp-agentic-review + security-reviewer.

Sources: Microsoft STRIDE (Spoofing, Tampering, Repudiation, Information Disclosure, Denial of Service, Elevation of Privilege) + OWASP Threat Modeling + OWASP Agentic Security (prompt injection, tool poisoning, excessive agency, data leakage, insecure plugin/MCP). See https://learn.microsoft.com/en-us/azure/security/develop/threat-modeling-tool-threats and https://owasp.org/www-project-threat-model/ .

When to use

  • New system/C4 design (architect output), new data flow (user → agent → tool → external), new trust boundary (skill composition, MCP addition, swarm handoff)
  • Before security-reviewer / owasp-agentic-review prioritizes mitigations
  • On architecture change — incremental mode re-runs only on deltas

Inputs

InputSourceRequired
Architecture diagramarchitect / C4/Mermaid (#369) — docs/architecture/*.md, Mermaid C4Context/flowchartYes (or textual description for small scopes)
AssetsCode + docs: skills, agents, MCP registry, secrets, data storesYes
Trust boundariesUser ↔ agent, agent ↔ MCP, agent ↔ external host, inter-agent, skill compositionYes
Data flowsaudit-capability.py shell/network/MCP/hooks surface, registry YAML, hook configsYes
ActorsUsers, agents (by persona), MCP servers, external servicesYes

Outputs

  • threat-model.md per references/threat-model-template.md: assets/boundaries/flows/actors → STRIDE + agentic threats table (severity, likelihood, impact, confidence, evidence) → mitigations → security acceptance criteria
  • Incremental diff when architecture discovery changed
  • Handoff: agentic-security-reviewer consumes findings for OWASP review; security-reviewer for app-layer vulns

STRIDE + agentic checklist

STRIDEMeaningAgentic extensionEvidence to check
S SpoofingImpersonate user/agent/MCPIdentity spoofing (AGNT01), prompt hierarchy violation (AGNT06)agents/*/AGENT.md identity, MCP auth.env, delegation tables
T TamperingModify data/code/instructions in transitTool poisoning (AGNT02), training-data/supply-chain poisoning (LLM03)mcp/registry/*.yaml tool descriptions with imperative injection, capabilities/upstream.lock digests
R RepudiationDeny action without audit trailMissing output-handshake, Swarm handoff without loggingoutput-handshake gates, ops/swarm-handoff audit
I Information disclosureLeak secrets/PII via tool outputData leakage (AGNT04), sensitive info disclosure (LLM06), model theft (LLM10)security.network_hosts, secret_storage, scans for ghp_, sk-, PII
D Denial of serviceExhaust tokens/computeModel DoS (LLM04) — unbounded skill context, swarm fan-outSkill token count, subagent fan-out limits, docs/research context audit
E Elevation of privilegeGain permissions beyond least privilegeExcessive agency (LLM08), permission creep (AGNT05), insecure plugin design (LLM07)approval.default, security.dangerous_permissions, composed skill chain

Map each finding to STRIDE letter(s) + agentic IDs (LLM01–LLM10, AGNT01–AGNT06) where applicable. One finding can map to multiple STRIDE categories (e.g., tool poisoning = T + E).

Workflow

  1. Discover scope: Identify assets + trust boundaries + data flows + actors from architecture diagram (or git diff HEAD deltas in incremental mode). Enumerate: skills (skills/**), agents (agents/**), MCP servers (mcp/registry/*.yaml + templates), hooks, subagents, external hosts.
  2. Enumerate threats: For each asset/flow/boundary, apply STRIDE + agentic checklist above. Ask per element: who can spoof? what can be tampered? what lacks repudiation? what leaks? what DoS? what elevates? Add agentic prompts: can tool description inject? can MCP exfiltrate? can agent over-act?
  3. Risk-rank: Score Impact (who/what compromised) × Likelihood (Low/Med/High) × Confidence (High/Med/Low) → Severity Critical/High/Medium/Low (or Blocking/Major/Minor). Cite evidence: file:line, registry YAML path, audit-capability.py output, diagram node.
  4. Mitigate + acceptance criteria: Per finding: mitigation (least-privilege, pin + provenance, sanitize, gate), residual risk, security acceptance criteria (testable, e.g., "tool descriptions contain no imperative injection via audit-capability.py clean"). Apply output-handshake before final artifact.
  5. Incremental review: On architecture change, diff threat-model.md previous version: keep unchanged findings, re-evaluate only changed assets/flows. For large systems, shard by trust boundary or swarm agent.

Risk table shape

#Asset / Flow (diagram ref)Trust boundarySTRIDEAgentic IDThreat (observation, file:line)Attack path (actor → boundary → asset)ImpactLikelihoodSeverityConfidenceMitigationResidualAcceptance criteriaEvidence

Sort by Severity (Critical → Low), then Likelihood. Include Residual risk after mitigation.

Incremental mode

  • Input: git diff + previous threat-model.md (from repo docs/security/threat-model*.md or prior PR).
  • Behavior: Carry forward findings for unchanged elements with status: unchanged; mark new/updated/removed for deltas; re-rank only affected rows.
  • Swarm-friendly: Split large architectures by trust boundary (e.g., user ↔ agent vs agent ↔ MCP) and run parallel subagents; merge tables, dedupe by asset+STRIDE.

Collaboration

NeedDelegate to
System design, C4/Mermaid, ADRsarchitect / platform-engineer (references/WORKSTATION_OPS.md) — provides Inputs architecture diagram
App/code vulns (OWASP Web)security-reviewer / agents/security-reviewer — consumes threat model for prioritized review
Agentic / LLM / MCP / supply-chainagentic-security-reviewer + agentic-security/owasp-agentic-review (LLM01-10 + AGNT01-06) — joint risk table
Full supply-chain surfaceagentic-security/supply-chain-audit
MCP config/impl depthagentic-security/mcp-audit
Output gateoutput-handshake
Swarm decompositionops/swarm-handoff

Sequence: architecture discovery → threat-modeling (this skill) → security review (owasp-agentic + security-reviewer) → security acceptance criteria.

Anti-patterns

  • Do not hallucinate STRIDE findings without evidence (cite file:line or registry path).
  • Do not collapse STRIDE into single category — one flow often maps to S+T+E.
  • Do not claim full coverage from automated checks alone — mark Not assessed (requires manual review) where needed (per #380 a11y analog).
  • Do not run full model on every file change — use incremental mode.

References

Keep looking

Skills are one crate of 610,677. Ordering is by how many stacks a row turns up in, so the top of any crate is what has actually been picked rather than what has the most stars.