Security: gate Anthropic-backed simulation analysis with auth and quotas

Summary

policyengine-api exposes an unauthenticated Anthropic-backed simulation analysis endpoint.

Severity

High

Impact

Attackers can burn LLM quota and persist arbitrary analysis records by repeatedly calling the endpoint.

Affected code

• policyengine_api/routes/simulation_analysis_routes.py:19-78
• policyengine_api/services/ai_analysis_service.py:45-77

Details

/simulation-analysis accepts public requests, and AIAnalysisService immediately initializes an Anthropic client with ANTHROPIC_API_KEY and streams a response for new prompts.

Expected behavior

LLM-backed analysis should require auth, rate limits, and ideally per-user quotas or explicit feature gating.

Suggested remediation

• Require authenticated callers
• Add rate limiting / quota enforcement
• Consider caching and internal-only generation paths for expensive prompts
• Add tests proving unauthenticated requests are denied

Start with policyengine_api/routes/simulation_analysis_routes.py:19-78 and policyengine_api/services/ai_analysis_service.py:45-77 to trace how unauthenticated requests reach Anthropic-backed analysis. Review the endpoint and service behavior, then add tests proving unauthenticated requests are denied and verify the intended rate-limit or quota enforcement. Done means public callers cannot repeatedly consume LLM quota or persist arbitrary analysis records.