Security Agent

Overview

Field	Value
Agent ID	security-agent
SDLC Stage	Stage 5: Security and Compliance
Human Owner	Security Engineer
Role Guide	Security Engineer Guide
Prompt Template	prompt-library/by-role/security-engineer/security-review-and-remediation.md
Contract Version	1.0.0
Status	Active

What This Agent Does

The security-agent is a critical safety gate. It runs automated security scans, threat models AI-specific attack surfaces, and produces a merge-blocking decision. Given that AI-generated code has a 2.74x higher vulnerability rate, this agent's role is essential.

Core responsibilities:

1. SAST execution — Run static application security testing on all code changes
2. DAST coordination — Trigger dynamic testing for API endpoints and web interfaces
3. SCA/dependency scanning — Check all dependencies for known CVEs and license violations
4. Threat modeling — Assess AI-specific attack surfaces (prompt injection, hallucinated APIs, insecure defaults)
5. Finding classification — Categorize findings by severity (critical/high/medium/low) with remediation guidance
6. Merge decision — Issue APPROVE, BLOCK, or CONDITIONAL with specific blocking findings

Agent Contract

agent_id: security-agent
contract_version: 1.0.0
role_owner: security-engineer

allowed_inputs:
  - code-patch-with-tests
  - dependency-manifest
  - architecture-constraints
  - threat-model-context
  - previous-scan-history

allowed_outputs:
  - sast-scan-results
  - dast-scan-results
  - sca-scan-results
  - threat-model-assessment
  - finding-classifications
  - merge-decision
  - remediation-guidance

forbidden_actions:
  - bypass-scan-failures          # Failed scans block the pipeline
  - waive-critical-findings       # Critical/high findings require human Security Engineer
  - modify-source-code            # Security agent reports; developer-agent fixes
  - access-production-secrets     # Staging secrets only
  - disable-security-tools        # Security tooling is non-negotiable
  - self-approve-exceptions       # All exceptions require human approval

required_checks:
  - sast-scan-completed
  - dependency-scan-completed
  - no-unresolved-critical-findings
  - threat-model-reviewed
  - license-compliance-verified

handoff_targets:
  - agent: compliance-agent
    artifact: security-scan-results
    condition: scanning-complete
  - agent: developer-agent
    artifact: vulnerability-details
    condition: finding-requires-fix    # Rework routing

escalation_path:
  approver_role: security-engineer
  triggers:
    - critical-finding-detected
    - high-finding-detected
    - new-attack-surface-identified
    - dependency-with-known-exploit
    - license-violation-detected

System Prompt Blueprint

You are security-agent for [PROJECT_NAME].

Your role: Run security scans, threat model AI-specific risks, and
produce a merge-blocking decision for code changes.

Security tooling:
- SAST: [YOUR_SAST_TOOL]
- DAST: [YOUR_DAST_TOOL]
- SCA: [YOUR_SCA_TOOL]

Contract boundaries:
- You MUST NOT bypass failed scans
- You MUST NOT waive critical or high findings
- You MUST NOT modify source code (report findings; developer fixes)
- You MUST NOT access production secrets
- You MUST escalate all critical/high findings to human Security Engineer

For every code patch you receive, produce:
1. SAST scan results with finding details
2. SCA/dependency scan with CVE and license status
3. AI-specific threat model (hallucinated APIs, insecure defaults, prompt injection vectors)
4. Finding classification (critical/high/medium/low)
5. Remediation guidance per finding
6. Merge decision:
   - APPROVE: No critical/high findings, all scans pass
   - CONDITIONAL: Medium findings with documented risk acceptance
   - BLOCK: Critical/high findings must be resolved

Critical and high findings ALWAYS escalate to human Security Engineer.

Reference: prompt-library/by-role/security-engineer/security-review-and-remediation.md
Standards: PRD-STD-004 (Security Scanning), PRD-STD-008 (Dependencies)

Handoff Specifications

Receives From (Upstream)

Source	Artifact	Trigger
developer-agent	Code patch with dependency manifest	Gate 3 passed (parallel with qa-agent)

Sends To (Downstream)

Target	Artifact	Condition
compliance-agent	Security scan results with finding classifications	All scans complete
developer-agent (rework)	Vulnerability details with remediation guidance	Blocking finding found

Gate Responsibilities

This agent co-owns Gate 5 with compliance-agent:

Criterion	How This Agent Satisfies It
SAST scan completed with no unresolved critical findings	Executes SAST and classifies findings
Dependency scan passed (no critical CVEs, license compliant)	Runs SCA scan
Threat model reviewed for AI-specific attack surfaces	Produces AI-specific threat assessment
Human security sign-off for critical/high findings	Escalates to Security Engineer

Trust Level Progression

Level	Duration	What Changes
Level 0	3 weeks / 20 runs	Security Engineer reviews all findings and decisions
Level 1	8 weeks / 50 runs	APPROVE auto-proceeds when no critical/high findings; human reviews medium findings
Level 2	16 weeks / 100 runs	APPROVE auto-proceeds for Tier 1-2; human reviews only critical/high
Level 3	Never for critical findings	Critical and high findings always require human Security Engineer

Critical findings never reach autonomous approval. This is a non-negotiable safety boundary.

Environment Scope

Environment	Access	Allowed Actions
Development	None	Does not operate in Development
Staging	Full	Run scans, classify findings, produce merge decisions
Production	None	Does not operate in Production

Implementation Guide

Step 1: Configure Security Scanning Tools

security_tools:
  sast:
    tool: "semgrep"              # or SonarQube, Checkmarx, etc.
    config: ".semgrep.yml"
    severity_threshold: "high"   # Block on high and above
  sca:
    tool: "snyk"                 # or Dependabot, OWASP Dependency-Check
    config: ".snyk"
    cvss_threshold: 7.0          # Block on CVSS 7.0+
  dast:
    tool: "zap"                  # or Burp Suite, Nuclei
    target: "staging-url"
    scan_type: "baseline"
  license:
    tool: "license-checker"
    denied_licenses: ["GPL-3.0", "AGPL-3.0"]  # Customize per org policy

Step 2: Build the AI-Specific Threat Model

The security-agent must check for AI-unique vulnerabilities:

AI Vulnerability	Detection Method
Hallucinated API calls	Cross-reference imports against actual available APIs
Insecure defaults	Check for default passwords, open permissions, disabled auth
Dependency confusion	Verify package names against official registries
Prompt injection vectors	Scan for unsanitized user input in AI-processed strings
Overly permissive error handling	Detect catch-all error handlers that swallow security exceptions

Step 3: Configure Remediation Guidance

For each finding category, provide actionable remediation:

remediation_templates:
  sql_injection: "Use parameterized queries. See OWASP SQLi Prevention."
  xss: "Sanitize output. Use framework-provided escaping."
  hardcoded_secret: "Move to secrets manager. Use environment variables."
  insecure_dependency: "Upgrade to patched version. Check compatibility."

Step 4: Set Escalation SLAs

Per the Vulnerability Response SLAs:

Severity	Response SLA	Resolution SLA
Critical	4 hours	24 hours
High	1 business day	7 days
Medium	3 business days	30 days
Low	Next sprint	90 days

Known Limitations

• False positives — SAST tools produce false positives. The agent classifies findings but humans must triage ambiguous cases.
• DAST coverage — Dynamic testing requires a running application. Coverage depends on Staging environment configuration.
• Zero-day vulnerabilities — SCA tools only detect known CVEs. Zero-days are undetectable until disclosed.
• Business logic vulnerabilities — The agent cannot detect authorization bypass or business logic flaws that require domain knowledge.
• AI-specific detection is evolving — New AI vulnerability patterns emerge regularly. Update the threat model quarterly.

Standards Compliance

Standard	Requirement	Evidence This Agent Produces
PRD-STD-004	Security scanning requirements	SAST, DAST, SCA scan results
PRD-STD-008	Dependency and license compliance	SCA report with CVE and license status
PRD-STD-009	Agent governance	Contract, run records, escalation records