Why compliance in AI development

The use of AI agents in software development introduces specific regulatory risks:

• Traceability: Who wrote each line of code? A human or an AI?
• Governance: How are AI agents configured and controlled?
• Transparency: Is the use of AI adequately documented?
• Security: Are there guardrails to prevent undesired behavior?
• Accountability: Is there human review before deploying AI-generated code?

licit evaluates these aspects against established regulatory frameworks.

EU AI Act (Regulation EU 2024/1689)

Scope

The EU AI Act is the first comprehensive regulatory framework for artificial intelligence. It entered into force in August 2024, with gradual enforcement until August 2027.

licit evaluates the articles relevant to development teams using AI agents:

Articles evaluated

Evaluator scoring

Each article has a dedicated evaluation method with numerical scoring. The score is converted to status with _score_to_status(score, compliant_at, partial_at):

The evaluator generates actionable recommendations with concrete licit commands (e.g., “Run: licit trace — to start tracking code provenance”).

FRIA — Fundamental Rights Impact Assessment

The FRIA (Fundamental Rights Impact Assessment) is mandatory for high-risk AI systems under Art. 27. licit generates an interactive FRIA in 5 steps with 16 questions and auto-detection of 8 fields:

1. System Description (5 questions): Purpose, AI technology, models, scope, human review.
2. Fundamental Rights Identification (4 questions): Personal data, employment, safety, discrimination.
3. Impact Assessment (3 questions): Risk level, maximum impact, detection speed.
4. Mitigation Measures (5 questions): Guardrails, scanning, testing, audit trail, additional measures.
5. Monitoring & Review (3 questions): Review frequency, responsible person, incident process.

Auto-detection: For fields like system_purpose, guardrails, security_scanning, testing, and audit_trail, licit infers the answer from the project’s ProjectContext and EvidenceBundle.

Command:

licit fria            # New interactive questionnaire
licit fria --auto     # Non-interactive mode (CI/CD)
licit fria --update   # Update existing FRIA

Generated files:

• .licit/fria-data.json — Raw data (JSON, reusable with --update)
• .licit/fria-report.md — Markdown report with Jinja2 template

Annex IV — Technical Documentation

Annex IV defines the technical documentation required for AI systems. licit generates this documentation by auto-populating it from 27 template variables extracted from:

• Project metadata (pyproject.toml, package.json)
• CI/CD configuration
• AI agent configurations
• Testing frameworks and security tools
• Provenance data (% AI code)
• Evidence of guardrails, quality gates, budget limits, FRIA, audit trail

6 auto-generated sections:

1. General Description — Purpose, AI components, languages, frameworks
2. Development Process — Version control, provenance, agent configs
3. Monitoring & Control — CI/CD, audit trail, changelog
4. Risk Management — Guardrails, quality gates, budget, oversight, FRIA
5. Testing & Validation — Test framework, security scanning
6. Changes & Lifecycle — Tracking mechanisms

Each section without evidence generates an actionable recommendation (e.g., “Run licit trace to begin tracking code provenance”).

Command:

licit annex-iv --organization "My Company" --product "My Product"

OWASP Agentic Top 10 (2025)

Scope

The OWASP Top 10 for Agentic AI Security identifies the top 10 security risks in applications using AI agents. licit evaluates the project’s posture against each risk with numerical scoring.

Status: Implemented since v0.5.0. Run with licit verify --framework owasp.

Risks evaluated

Evaluator scoring

Each risk has a dedicated evaluation method with numerical scoring. The score is converted to status with _score_to_status(score, compliant_at, partial_at):

ASI08 and ASI09 use compliant_at=2 because they have fewer available signals. The evaluator generates actionable recommendations with concrete tools (e.g., “Add AI-specific security scanning: vigil detects prompt injection patterns”).

Design note: The evaluator measures the presence of security tools, not their findings. A project with vigil installed but 50 critical findings gets the same score as one with 0 findings. Findings are relevant for the gap analyzer (Phase 6).

Mapping to evidence

Each OWASP risk maps to collectible evidence from ProjectContext and EvidenceBundle:

ASI01 (Excessive Agency)
  ├── ev.has_guardrails + ev.guardrail_count
  ├── ev.has_quality_gates + ev.quality_gate_count
  ├── ev.has_budget_limits
  └── ctx.agent_configs

ASI02 (Prompt Injection)
  ├── ctx.security.has_vigil (+2 — AI-specific scanning)
  ├── ev.has_guardrails
  └── ev.has_human_review_gate

ASI04 (Logging & Monitoring)
  ├── ctx.git_initialized + ctx.total_commits
  ├── ev.has_audit_trail + ev.audit_entry_count (+2)
  ├── ev.has_provenance + ev.provenance_stats
  └── ev.has_otel

ASI06 (Human Oversight)
  ├── ev.has_human_review_gate (+2 — critical control)
  ├── ev.has_dry_run
  ├── ev.has_quality_gates
  └── ev.has_rollback

ASI08 (Unbounded Resources)
  ├── ev.has_budget_limits (+2 — direct control)
  └── ev.has_quality_gates

Command:

licit verify --framework owasp   # Evaluate OWASP only
licit verify --framework all     # Evaluate EU AI Act + OWASP

How licit evaluates compliance

Evaluation process

1. Detect     → ProjectDetector analyzes the project
2. Collect    → EvidenceCollector gathers evidence
3. Evaluate   → Evaluators apply framework requirements
4. Classify   → Each requirement: compliant / partial / non-compliant / n/a
5. Report     → Report with evidence, gaps, and recommendations

Evidence sources

Provenance evidence (licit trace) directly feeds the transparency (Art. 13) and traceability (Art. 10) articles of the EU AI Act. The config changelog (licit changelog) feeds the transparency (Art. 13) and deployer obligations (Art. 26) articles. Both feed the monitoring controls (ASI-06, ASI-10) of the OWASP Agentic Top 10.

Compliance levels

Compliance reports

Available formats

Report structure

# Compliance Report — My Project
Generated: 2026-03-10

## Summary
- EU AI Act: 72% compliant (13/18 controls)
- OWASP Agentic: 60% compliant (6/10 controls)

## EU AI Act
### Article 9 — Risk Management
Status: PARTIAL
Evidence: FRIA exists but incomplete
Recommendation: Complete FRIA sections 3-5

### Article 14 — Human Oversight
Status: COMPLIANT
Evidence: GitHub Actions requires approval for deployment
...

## Gaps
| Priority | Requirement | Gap | Effort |
|---|---|---|---|
| 1 | ART-9-1 | No risk assessment | Medium |
| 2 | ASI-01 | No guardrails | Low |

CI/CD Gate

licit can act as a compliance gate in CI/CD pipelines:

# .github/workflows/compliance.yml
name: Compliance Check
on: [push, pull_request]

jobs:
  compliance:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
        with:
          fetch-depth: 0   # Required for git analysis

      - uses: actions/setup-python@v5
        with:
          python-version: "3.12"

      - name: Install licit
        run: pip install licit-ai-cli

      - name: Run compliance check
        run: licit verify
        # Exit 0 = pass, Exit 1 = fail, Exit 2 = warnings

Exit codes:

Future frameworks (V1+)

licit is designed to support additional frameworks:

The frameworks/ architecture allows adding new frameworks by implementing an evaluator with the corresponding Protocol interface.