🔒 Agent 5

Security Agent

The Security Agent performs a tiered, 20-domain security review of the built codebase. It activates after the Critic gate (!security) — no project advances to Verifier without a completed security review. Critical and high-severity findings block the pipeline until resolved.

🚨

Security review is non-optional

The Security Agent cannot be skipped, bypassed, or fast-tracked. Any attempt to jump directly from Builder to Verifier without !security is refused by the Orchestrator.

Pre-Populated Build Phase Findings

At activation, the Security Agent immediately reads two artifact files produced during the build phase before performing its own domain review:

File	Produced by	Contents
`SAST_FINDINGS.jsonl`	`sast_scan.py` (runs after every write)	Static analysis findings: rule ID, severity, file, line, message
`SECRET_BLOCKS.jsonl`	`secret_scan.py` (runs pre-write)	Detected hardcoded secrets that were blocked during build — file, pattern type, line

SAST Findings Handling

Every pre-identified finding in SAST_FINDINGS.jsonl must be explicitly addressed in SECURITY_REPORT.md. The Security Agent may not silently skip any finding:

Confirmed — finding is valid; remediation is required before !verify
Expanded — finding is valid and the agent identifies additional related issues
Dismissed — finding is a false positive; documented rationale required (e.g., "test file, not production code")

Secret Blocks Handling

For each entry in SECRET_BLOCKS.jsonl, the Security Agent verifies:

The hardcoded value was actually replaced with an environment variable reference
The environment variable is documented in .env.example

🚨

Remaining hardcoded secrets = automatic BLOCK (Critical)

Any secret listed in SECRET_BLOCKS.jsonl that is still hardcoded in the codebase at Security review time is classified as CRITICAL and blocks the pipeline unconditionally. This finding cannot be dismissed or accepted — it must be fixed.

ℹ️

Model-adaptive behavior

The Security Agent adjusts its report format and tooling assumptions based on the active model family. Claude models produce full SECURITY_REPORT.md with all sections. Kimi and GPT-family models use a condensed format. Ollama local models skip dependency intelligence scans that require network access.

Complexity-Tiered Review

The number of domains reviewed depends on the project's complexity tier from Discovery:

Tier	Domains reviewed	Typical scope
Simple	5 (LOW severity domains)	Static sites, single-user tools, no auth, no external data
Standard	HIGH and CRITICAL domains	Auth, database, API integrations, multi-user apps
Complex	All 20 domains	Payments, multi-tenant, ML/AI, compliance requirements

20 Security Domains

D1 Authentication & Session Management

D2 Authorization & Access Control

D3 Input Validation & Sanitization

D4 SQL & NoSQL Injection

D5 Cross-Site Scripting (XSS)

D6 Cross-Site Request Forgery (CSRF)

D7 Secrets & Key Management

D8 Dependency & Supply Chain

D9 API Security & Rate Limiting

D10 Data Encryption (at rest + in transit)

D11 File Upload & Storage Security

D12 Error Handling & Information Leakage

D13 Logging & Audit Trail

D14 Payment & Financial Security

D15 Mobile-Specific Threats

D16 AI/LLM Prompt Injection

D17 Infrastructure & Deployment Security

D18 Privacy & Data Residency (GDPR/CCPA)

D19 Multi-Tenancy Isolation

D20 Business Logic & Abuse Prevention

Finding Severity Levels

Severity	Description	Pipeline impact
CRITICAL	Exploitable vulnerability that could result in data breach, account takeover, or system compromise	Hard block — pipeline cannot advance until resolved
HIGH	Significant risk that requires a fix before production deployment	Soft block — must be resolved or formally accepted before `!verify`
MEDIUM	Risk that should be addressed but doesn't block shipping an MVP	Logged in report; tracked in HANDOFF.json v5
LOW	Hardening opportunity — defense in depth improvement	Logged for awareness; no pipeline impact

Report Format

The Security Agent writes artifacts/security/SECURITY_REPORT.md with:

Executive Summary with overall risk rating
Per-domain findings (domain, severity, finding, evidence, remediation)
Critical findings list (pipeline blockers)
Dependency vulnerability scan results
Secrets audit (hardcoded values, missing env vars)
Platform-specific threat model notes
Recommended remediations ordered by severity
HANDOFF.json v5 update

Security Commands

Command	Behavior
!security	Activate the Security Agent; begin tiered 20-domain review
!security domain [N]	Run a specific domain review only (useful for re-review after fixes)
!security accept [ID]	Formally accept a HIGH or MEDIUM finding with documented rationale (adds to HANDOFF.json v5 as accepted_risk)