Security Compliance

Security & Compliance

What security teams need to approve AI coding assistant adoption

---

Overview

Security teams block AI coding assistant adoption because:

1. No governance - AI generates code without review
2. No audit trail - Can't trace what AI did
3. Security blind spots - AI may introduce vulnerabilities
4. Compliance gaps - No evidence of controls

AssemblyZero addresses each concern with concrete mechanisms.

---

Governance Framework

Multi-Layer Review

┌─────────────────────────────────────────────────┐
│                 HUMAN REVIEW                     │
│         (Final approval before merge)            │
├─────────────────────────────────────────────────┤
│              GEMINI REVIEW GATE                  │
│  (AI reviews AI before human sees it)            │
├─────────────────────────────────────────────────┤
│              CLAUDE GENERATION                   │
│        (AI generates code/designs)               │
└─────────────────────────────────────────────────┘

Every AI-generated artifact passes through:

1. Gemini verification - Different model reviews for issues
2. Human approval - Final judgment before merge

Enforced Gates

Three mandatory checkpoints:

Gate	What's Reviewed	Evidence
LLD Review	Design document	Gemini verdict + reasoning
Implementation Review	Code + tests	Gemini verdict + reasoning
Report Generation	Auto-documentation	Audit trail

Gates cannot be skipped (prompt-based now, state machine in roadmap).

---

Security Audits

OWASP Top 10 Coverage

AssemblyZero includes security audits for OWASP vulnerabilities:

OWASP Category	Audit	Status
A01: Broken Access Control	Access control audit	✓
A02: Cryptographic Failures	Crypto audit	✓
A03: Injection	Injection audit	✓
A04: Insecure Design	Design review (LLD gate)	✓
A05: Security Misconfiguration	Config audit	✓
A06: Vulnerable Components	Dependency audit	✓
A07: Auth Failures	Auth audit	✓
A08: Software Integrity Failures	Integrity audit	✓
A09: Logging Failures	Logging audit	✓
A10: SSRF	SSRF audit	✓

Audit Philosophy

Adversarial, not confirmatory.

WRONG: "Check that code follows security best practices" ✗
RIGHT: "Find any injection vulnerabilities in this code" ✓

Audits are designed to find problems, not confirm compliance.

Running Audits

# Run full security audit
/audit --type security

# Run specific OWASP audit
/audit --type owasp-injection

# Audit with auto-fix suggestions
/audit --type security --fix

---

Privacy Compliance

GDPR Considerations

Requirement	AssemblyZero Control
Data minimization	Agents access only necessary files
Purpose limitation	Session scoped to specific task
Storage limitation	Session data retention policies
Audit trail	Full session transcript logging

Data Handling

• No PII in prompts - Agents instructed to avoid PII
• Local processing - File reads are local, not sent upstream
• Session isolation - Each session is independent
• Transcript retention - Configurable retention period

---

AI Safety (NIST AI RMF)

Key Principles

NIST Category	AssemblyZero Implementation
Validity	Gemini verification of outputs
Safety	Destructive command blocks
Security	OWASP audits, secret detection
Accountability	Full audit trail
Transparency	Visible reasoning in transcripts
Explainability	Reports document decisions
Privacy	Data minimization, local processing

Destructive Command Protection

Certain commands are blocked regardless of context:

ALWAYS BLOCKED (catastrophic risk):
├── dd if=...          # Disk operations
├── mkfs               # Filesystem creation
├── shred              # Secure delete
└── format             # Format disk

PATH-SCOPED (allowed only in Projects):
├── rm, rm -r, rm -rf  # File deletion
├── git reset --hard   # Discard changes
└── git push --force   # Overwrite history

Model Verification

Agents must use approved models:

Model	Approved For
claude-opus-4	Complex reasoning, architecture
claude-sonnet-4	General development tasks
gemini-3-pro	Review and verification

Silent downgrades are detected and rejected.

---

Audit Trail

What's Logged

Every agent session creates:

Session: 2026-01-21-001
├── Transcript (full conversation)
├── Tool calls (all commands executed)
├── File changes (diffs)
├── Gemini reviews (verdicts + reasoning)
├── Reports (implementation, tests)
└── Metadata (timestamps, models used)

Log Format

{"timestamp": "2026-01-21T14:32:00Z", "type": "tool_call", "tool": "Bash", "command": "git status"}
{"timestamp": "2026-01-21T14:32:01Z", "type": "tool_result", "output": "On branch feature..."}
{"timestamp": "2026-01-21T14:33:00Z", "type": "gemini_review", "verdict": "APPROVE", "model": "gemini-3-pro"}

Retention

Default retention: 90 days Configurable in: ~/.agentos/config.json

---

Compliance Evidence

For SOC 2

Control	Evidence
Access Control	Single-user identity model
Change Management	PR workflow, Gemini gates
Audit Logging	Session transcripts
Risk Assessment	34 audits covering major risks

For ISO 27001

Control Area	Evidence
A.12 Operations Security	Destructive command blocks
A.14 System Acquisition	Gemini verification gates
A.16 Incident Management	Audit logs for investigation
A.18 Compliance	OWASP, GDPR audits

For HIPAA (if applicable)

Requirement	Evidence
Access Controls	Single-user model, session isolation
Audit Controls	Full session logging
Integrity Controls	Gemini verification
Transmission Security	Local processing, no PII in prompts

---

34 Audits Catalog

By Category

Category	Count	Examples
Security (OWASP)	10	Injection, XSS, auth
Privacy (GDPR)	6	Data handling, consent
AI Safety (NIST)	5	Model verification, bias
Code Quality	8	Complexity, duplication
Documentation	5	Completeness, accuracy

Audit Index

Full catalog: /audit --list

Security Audits:
├── 0801-injection-audit
├── 0802-xss-audit
├── 0803-auth-audit
├── 0804-access-control-audit
├── 0805-crypto-audit
├── 0806-config-audit
├── 0807-dependency-audit
├── 0808-ssrf-audit
├── 0809-logging-audit
└── 0810-integrity-audit

Privacy Audits:
├── 0811-pii-detection
├── 0812-consent-audit
├── 0813-data-retention-audit
├── 0814-data-minimization-audit
├── 0815-cross-border-audit
└── 0816-dsar-readiness-audit

AI Safety Audits:
├── 0817-model-verification
├── 0818-bias-detection
├── 0819-hallucination-check
├── 0820-prompt-injection-audit
└── 0821-output-validation-audit

---

For Security Team Reviews

Approval Checklist

Before approving AI coding assistant adoption:

• Multi-model verification enabled (Claude + Gemini)
• All three gates enforced (LLD, Implementation, Report)
• OWASP audits passing (90%+ coverage)
• Destructive commands blocked outside Projects
• Session logging enabled
• Retention policy defined
• Model verification active (no silent downgrades)

Risk Assessment

Risk	Likelihood	Impact	Mitigation
AI introduces vulnerability	Medium	High	Gemini security review
Unauthorized code changes	Low	High	PR workflow required
Sensitive data exposure	Medium	High	PII detection audit
Silent model degradation	Low	Medium	Model verification
Audit trail gaps	Low	High	Session logging

Residual Risk

With AssemblyZero controls:

• Reduced but not eliminated risk of AI-introduced bugs
• Mitigated risk of security vulnerabilities
• Maintained audit trail for incident response
• Preserved human judgment for final approval

---

Related Pages

• Governance Gates - Gate implementation details
• Gemini Verification - Multi-model architecture
• Measuring Productivity - Audit metrics
• Multi-Agent Orchestration - Overall architecture