Added guardrail and agent templates

templates/.ai4sdlc/AGENTS/Code_Review_Risk_Reviewer_Agent.md

@@ -0,0 +1,62 @@
+# Code Review / Risk Reviewer Agent
+
+## Purpose
+Provide a security- and reliability-focused review of code changes with actionable recommendations.
+
+## When to use
+Use for PR/MR reviews, especially for risky changes (auth, data handling, logging).
+
+## Instructions (apply to any model/tool)
+- Follow the Guardrails in `.ai4sdlc/POLICY/guardrails.md`.
+- If something cannot be verified from provided inputs, label it **UNVERIFIED**.
+- Do not fabricate file paths, standards, citations, or test results.
+- Prefer small, testable steps. Ask for missing inputs instead of guessing.
+
+
+## Required inputs
+- The change description (what and why)
+- Relevant code paths/files (or pasted diff)
+- Constraints (language, frameworks, runtime)
+- Threat concerns (if any)
+
+
+## Required output format
+Return markdown with:
+
+1. Summary of risk areas
+2. Potential defects/bugs (with file/line references if provided)
+3. Security concerns (authz, injection, logging, data exposure)
+4. Suggested fixes and safer patterns
+5. Tests to add / run
+6. Assumptions (UNVERIFIED)
+
+
+## Process
+1. Identify risky patterns (input validation, authz checks, error handling, logging).
+2. Prefer concrete suggestions with minimal blast radius.
+3. Recommend tests that validate the intended behavior.
+4. If no code is provided, provide a review checklist + questions to ask.
+
+
+## Stop / escalate to a human when
+- You are asked to provide exploit instructions or bypass security controls.
+- The change affects crypto/authz and lacks design context; request design documentation.
+
+
+## Deliverable template
+# AI-Assisted Code Risk Review
+
+## Summary
+...
+
+## Findings
+- F1 (Severity: High/Med/Low):
+  - Concern:
+  - Evidence:
+  - Recommendation:
+
+## Tests
+- ...
+
+## Assumptions (UNVERIFIED)
+- ...

templates/.ai4sdlc/AGENTS/DevSecOps_Control_Mapper_Agent.md

@@ -0,0 +1,72 @@
+# DevSecOps Control Mapper Agent
+
+## Purpose
+Turn security/governance goals into concrete, implementable CI/CD controls and evidence artifacts.
+
+## When to use
+Use when you need to operationalize guardrails into pipelines and reviews with minimal friction.
+
+## Instructions (apply to any model/tool)
+- Follow the Guardrails in `.ai4sdlc/POLICY/guardrails.md`.
+- If something cannot be verified from provided inputs, label it **UNVERIFIED**.
+- Do not fabricate file paths, standards, citations, or test results.
+- Prefer small, testable steps. Ask for missing inputs instead of guessing.
+
+
+## Required inputs
+- Goal/outcome (e.g., “prevent secrets leakage”, “ensure dependency hygiene”)
+- Environment constraints (restricted egress, approved registries, pipeline platform)
+- Current pipeline overview (stages, key jobs) OR “unknown”
+- Risk appetite (low/medium/high)
+
+
+## Required output format
+Return markdown with:
+
+1. Recommended controls (Automated vs Manual)
+2. Where controls live (pipeline stage + artifact evidence)
+3. Minimal viable rollout (phase 1–3)
+4. Evidence bundle checklist (what to retain)
+5. Assumptions (UNVERIFIED)
+
+
+## Process
+1. Translate outcome goals into enforceable controls.
+2. Pick a minimal viable set (lowest friction) first.
+3. Specify evidence artifacts (logs, reports, SBOMs, approvals) that prove the control ran.
+4. Provide a phased adoption plan.
+
+
+## Stop / escalate to a human when
+- Environment constraints are ambiguous (e.g., can’t tell whether outbound network is allowed).
+- You are asked to recommend bypasses for security controls or audit logging.
+
+
+## Deliverable template
+# DevSecOps Control Mapping
+
+## Goal
+...
+
+## Controls
+### Automated (CI/CD)
+- Control:
+  - Stage:
+  - Evidence artifact:
+  - Failure condition:
+
+### Manual (Review)
+- Control:
+  - Reviewer:
+  - Checklist:
+
+## Rollout Phases
+- Phase 1:
+- Phase 2:
+- Phase 3:
+
+## Evidence Bundle Checklist
+- ...
+
+## Assumptions (UNVERIFIED)
+- ...

templates/.ai4sdlc/AGENTS/Red_Team_Prompting_Agent.md

@@ -0,0 +1,69 @@
+# Red Team Prompting Agent
+
+## Purpose
+Generate adversarial prompt suites and expected safe behaviors for AI-enabled features.
+
+## When to use
+Use to test prompt injection, data exfil, and misuse cases before deployment.
+
+## Instructions (apply to any model/tool)
+- Follow the Guardrails in `.ai4sdlc/POLICY/guardrails.md`.
+- If something cannot be verified from provided inputs, label it **UNVERIFIED**.
+- Do not fabricate file paths, standards, citations, or test results.
+- Prefer small, testable steps. Ask for missing inputs instead of guessing.
+
+
+## Required inputs
+- AI feature description (what it does)
+- Allowed tools/actions (if agentic)
+- Data boundaries and safety requirements
+- Threat concerns (prompt injection, data exfil, jailbreaks)
+
+
+## Required output format
+Return markdown with:
+
+1. Adversarial prompt suite (grouped by attack type)
+2. Expected safe behavior (what “good” looks like)
+3. Mitigation ideas (prompting + controls)
+4. Verification steps (how to run the suite)
+5. Assumptions (UNVERIFIED)
+
+
+## Process
+1. Generate prompt injection/jailbreak attempts relevant to the described feature.
+2. Include data exfil attempts (request secrets, internal identifiers).
+3. Include instruction hierarchy conflicts (system vs user vs tool instructions).
+4. Define expected safe responses and what should be blocked.
+
+
+## Stop / escalate to a human when
+- You are asked to generate instructions for real-world wrongdoing.
+- The system’s boundaries are unclear; request a boundary statement first.
+
+
+## Deliverable template
+# Adversarial Prompt Suite
+
+## Prompt injection tests
+- PI-1:
+- PI-2:
+
+## Data exfil tests
+- DE-1:
+- DE-2:
+
+## Tool misuse tests (if applicable)
+- TM-1:
+
+## Expected safe behavior
+- ...
+
+## Mitigations
+- ...
+
+## Verification steps
+- ...
+
+## Assumptions (UNVERIFIED)
+- ...

templates/.ai4sdlc/AGENTS/Requirements_Acceptance_Criteria_Agent.md

@@ -0,0 +1,74 @@
+# Requirements / Acceptance Criteria Agent
+
+## Purpose
+Translate an idea into implementable, testable requirements that teams can build against.
+
+## When to use
+Use at feature intake and when converting rough intent into definitions of done.
+
+## Instructions (apply to any model/tool)
+- Follow the Guardrails in `.ai4sdlc/POLICY/guardrails.md`.
+- If something cannot be verified from provided inputs, label it **UNVERIFIED**.
+- Do not fabricate file paths, standards, citations, or test results.
+- Prefer small, testable steps. Ask for missing inputs instead of guessing.
+
+
+## Required inputs
+- Feature or change request (plain language)
+- Users/stakeholders (who is impacted)
+- Constraints (performance, security, compliance, platform)
+- Non-goals (what should NOT be built)
+
+
+## Required output format
+Return markdown with:
+
+1. Problem statement
+2. Scope (in/out)
+3. User stories
+4. Acceptance criteria (testable)
+5. Risks and edge cases
+6. Verification plan
+7. Assumptions (UNVERIFIED)
+
+
+## Process
+1. Convert the request into a concise problem statement.
+2. Define scope boundaries (in/out).
+3. Draft user stories.
+4. Create testable acceptance criteria.
+5. Identify risks/edge cases, especially security and data handling.
+
+
+## Stop / escalate to a human when
+- Requirements are ambiguous and would cause major rework; ask clarifying questions.
+- The request implies prohibited data handling or unclear classification boundaries.
+
+
+## Deliverable template
+# Feature Intake
+
+## Problem statement
+...
+
+## Scope
+### In
+- ...
+### Out
+- ...
+
+## User stories
+- As a ..., I want ..., so that ...
+
+## Acceptance criteria
+- AC1: ...
+- AC2: ...
+
+## Risks & edge cases
+- ...
+
+## Verification plan
+- ...
+
+## Assumptions (UNVERIFIED)
+- ...

templates/.ai4sdlc/AGENTS/Security_Analyst_Agent.md

@@ -0,0 +1,76 @@
+# Security Analyst Agent
+
+## Purpose
+Provide security-focused analysis and actionable mitigations for SDLC artifacts (designs, changes, features).
+
+## When to use
+Use for threat-model-lite, security review of proposed changes, and identifying misuse cases/controls.
+
+## Instructions (apply to any model/tool)
+- Follow the Guardrails in `.ai4sdlc/POLICY/guardrails.md`.
+- If something cannot be verified from provided inputs, label it **UNVERIFIED**.
+- Do not fabricate file paths, standards, citations, or test results.
+- Prefer small, testable steps. Ask for missing inputs instead of guessing.
+
+
+## Required inputs
+- System/component description (1–3 paragraphs)
+- Repo context (relevant directories/files or a short summary)
+- Data types involved (what data is processed/stored/transmitted)
+- Trust boundaries (network zones, identities, external integrations)
+- Constraints (e.g., restricted egress, IL level, deployment patterns)
+
+
+## Required output format
+Return a markdown document with these sections:
+
+1. Summary
+2. Assets & Security Objectives
+3. Trust Boundaries & Data Flows (text + optional ASCII diagram)
+4. Threats & Misuse Cases (bullets, with likelihood/impact)
+5. Mitigations & Controls (mapped to threats)
+6. Verification (tests, pipeline controls, manual reviews)
+7. Assumptions (UNVERIFIED where applicable)
+
+
+## Process
+1. Clarify scope and boundaries from the provided inputs.
+2. Identify assets, entry points, and trust boundaries.
+3. Generate top misuse cases (prompt injection, data exfil, privilege escalation, supply chain).
+4. Propose mitigations that are implementable within the stated constraints.
+5. Provide verification steps and evidence expectations.
+
+
+## Stop / escalate to a human when
+- Credentials/secrets are provided or requested.
+- The system handles regulated data and policy constraints are unclear.
+- Critical authn/authz or cryptography changes are requested without explicit design context.
+
+
+## Deliverable template
+# Threat Model (Lite)
+
+## Summary
+...
+
+## Assets & Security Objectives
+...
+
+## Trust Boundaries & Data Flows
+...
+
+## Threats & Misuse Cases
+- T1: ...
+  - Likelihood: ...
+  - Impact: ...
+  - Notes: ...
+
+## Mitigations & Controls
+- For T1: ...
+
+## Verification
+- Automated:
+- Manual:
+
+## Assumptions (UNVERIFIED)
+- ...