🛡️ Security Ecosystem

This skill is the entry point of its ecosystem. It chooses a workflow, checks which member skills are installed, runs them in order with explicit handoffs, and degrades gracefully when a member is missing. The workflow map lives in references/workflows.json; the handoff format lives in references/handoff-contract.md.

Catalog

Domain: Security — Audit skills before deployment, guard agent runtimes against injection and exfiltration, and learn from incidents.

Philosophy: No skill or message is trusted by default. Audit before deploy, guard at runtime, and feed every incident back into the rules. Security verdicts come from deterministic rules and human review, not from a single model judgment.

Router Protocol

Follow this protocol whenever this ecosystem is invoked.

1. Match. Compare the request with each workflow's When phrases. Pick one workflow, or route straight to a single member skill when the request is narrow. If two workflows fit equally, ask one question to choose. If the request belongs to another domain, use the ecosystem index in @ecosystem-orchestration.
2. Preflight. Find out which step skills are installed: check the skills your host lists, or run python <skill-dir>/scripts/install_all.py --status --workflow <id> (<skill-dir> is the folder containing this SKILL.md). Show a short plan: step, skill, installed or missing, condition, expected artifact.
3. Missing skills never block. Offer once: (a) install them — install_all.py --workflow <id> prints the commands and only runs them with --execute after the user agrees; or (b) continue in fallback mode — perform the step yourself using its Does and Done when columns and label that output fallback.
4. Conditions. Skip optional steps whose condition does not hold, and say which steps you skipped and why.
5. Execute. Run steps in order. Steps that share a parallel group are independent and may run concurrently (@parallel-planner). Pass artifacts as described in references/handoff-contract.md. Check each step's Done when before moving on.
6. Gates. Before any destructive, irreversible, or production-impacting action, stop at @checkpoint-guardian, or ask for explicit confirmation when it is not installed.
7. Failures. When a step misses its Done when, hand off to @error-recovery. Without it, retry once with a changed approach, then report the blocker with evidence instead of guessing.
8. State. Keep a run log (workflow id, step, status, artifact). Record it with @memory-ledger when installed, and write a @session-transfer handoff if the run will continue in another session.
9. Finish. Summarize artifacts, skipped steps, open risks, and the most relevant next workflow.

Skip this protocol when the user names one specific skill (invoke it directly) or asks a quick factual question that needs no workflow.

Workflows

1. Skill Audit & Trust Flow (sec-skill-audit)

When: “is this skill safe”; “audit a new skill”; “skill trust score”

2. Runtime Protection Flow (sec-runtime-guard)

When: “add a security layer”; “prevent prompt injection”; “protect the agent at runtime”

3. Full Security Lifecycle (sec-full-lifecycle)

When: “run the security workflow”; “security orchestrator”; “audit and guard everything”

Direct handoffs

• @adaptive-guard → @security-auditor: An incident requires re-auditing the affected skill.
• @security-auditor → @skill-comparator: A failing skill should be replaced by a safer alternative.

Ecosystem Map

PRE-DEPLOYMENT         RUNTIME               LEARNING
──────────────         ─────────────         ──────────
security-auditor    →    adaptive-guard   →    Rule engine
  │                      │                    update
  │ Audit                │ K0-K4               │
  │ Trust Score          │ Gradual filter      │ async
  │ Guard rules          │ LLM Judge           │
  ▼                      ▼                    ▼
security-report.md   guard-decision.json   learned-rules.md
trust-score.md       incident-report.md    guard-metrics.md

Layer legend: K0–K4 are adaptive-guard's layers (K = katman, layer): K0 hash cache, K1 rule engine, K2 ML classifier, K3 LLM judge, K4 human approval.

Data Flow

security-auditor audit output
  → trust-score.md (used by adaptive-guard as trust threshold)
  → security-report.md (added to adaptive-guard K1 rules)

adaptive-guard attack detection
  → incident-report.md (appended to security-auditor's next audit)
  → learned-rules.md (feedback loop to K1 static rules)

Supported Platforms

Expansion Roadmap

Skills to be added: