The CouncilForge is at v1.0.0. Security fixes are applied to the latest main branch and released as patch versions. We do not maintain backport branches for older releases at this time.
| Version | Supported |
|---|---|
| 1.0.x | ✅ Latest main |
| < 1.0 | ❌ Not supported |
This project takes API key security seriously. The following guarantees are enforced throughout the codebase:
- API keys you configure through the Providers tab are stored in the local SQLite database in the
ProviderConfigtable. - The masked representation (e.g.
sk-...abc4) is stored inapiKeyMaskedand shown in the UI. - The actual key value is stored in
apiKeyEnc.⚠️ For local / single-user deployments only. TheapiKeyEncfield is currently stored as the raw value. Before exposing this app to multiple users or to the internet, you must:- Add real envelope encryption (e.g. AES-GCM with a key derived from an environment variable or a secrets manager).
- Put the deployment behind authentication (NextAuth is a dependency and ready to be wired up).
- Serve over HTTPS and set secure cookies.
- API keys are never written to the application log, the dev/server log files (
dev.log,server.log), or the observability event log. - The provider router records
providerId,providerName,reason,fallback, andmodel— never the key itself. - Audit log entries record
action: "provider_key_added"/"provider_key_removed"with the provider ID only, never the key.
- The policy engine runs
checkOutputPolicies()on every generated document and rejects output that looks like a leaked secret (AWS keys, OpenAI keys, JWTs, private keys, connection strings). - Generated documents, exported files (JSON / CSV / Markdown / issue-tracker format), and the decision journal are all scanned for credential patterns before being persisted or returned.
- If a leak is detected, the relevant section is redacted and a
policy_violationnotification is raised.
- Keys are only sent to the provider they belong to (e.g. an OpenRouter key is sent only to
openrouter.aias anAuthorizationheader). - Keys are never included in prompts, document content, repository fetches, or anything sent to GitHub.
- The
freemode (z-ai-web-dev-sdk) requires no key and never transmits your other keys anywhere.
- In
localmode, the configuredbaseURL(e.g.http://localhost:11434/v1for Ollama) is the only endpoint contacted. No traffic leaves your machine.
Please do not report security vulnerabilities through public GitHub issues.
Instead, use GitHub's private vulnerability reporting (Security → Report a vulnerability). This keeps the report private until a fix is coordinated. Include:
- A description of the vulnerability and its impact.
- Steps to reproduce, or a proof-of-concept.
- Affected versions (default to
mainif unsure). - Any suggested mitigations.
You should receive an acknowledgement within 72 hours. We'll work with you to understand the scope, develop a fix, and coordinate disclosure. Responsible reporters will be credited in the release notes (unless they prefer to remain anonymous).
Vulnerabilities in this repository are in scope, including:
- API key leakage (logs, documents, exports, observability).
- Authentication / authorization bypass (once auth is wired up).
- SQL injection or Prisma query issues.
- SSRF in the GitHub fetcher.
- XSS or CSRF in the workspace UI.
- Prompt injection that exfiltrates keys or secrets into documents.
- Vulnerabilities in upstream dependencies (report to the upstream maintainer; we'll bump versions once a fix is available).
- Issues in providers themselves (report to the provider).
- Theoretical attacks requiring physical access to the server.
- Self-XSS requiring the victim to paste the payload themselves.
- Rate limiting or denial-of-service on the dev server (it's not a production target).
The provider registry is bootstrapped from cheahjs/free-llm-api-resources. Treat that source as trusted but unverified — review any provider's terms of service before sending it data, and never send regulated data (PII, PHI, payment data) to a provider you haven't vetted.
- Free tiers often log prompts for abuse detection. Do not send sensitive data to free-tier providers.
- Rate limits are aggressive; expect fallbacks and partial failures.
- Provider availability is not guaranteed — the router's fallback chain handles this, but plan for it.
- Your keys are stored locally (see "Storage" above). Rotate them regularly.
- Use provider-specific features where available: OpenRouter's privacy mode, Together AI's SOC 2 controls, Google AI Studio's data controls.
- Set spending caps at the provider to avoid runaway costs during long pipeline runs.
- Verify your local endpoint (Ollama, LM Studio, vLLM) is bound to
localhostand not exposed to the network. - Local models can still leak data if the endpoint is shared — keep it single-tenant.
src/lib/github.tsuses the public GitHub REST API (no auth) with aggressive rate-limit handling.- It fetches repository metadata, README, and the file tree. It does not clone or execute code from the repository.
- Do not feed it URLs to private repositories — those will fail silently.
If you intend to deploy this beyond a local single-user dev environment:
- Enable real envelope encryption for
ProviderConfig.apiKeyEnc. - Wire up NextAuth with a real session store (the dependency is already installed).
- Serve over HTTPS with HSTS.
- Set
DATABASE_URLto a managed database (PostgreSQL works — change the Prismadatasourceprovider). - Restrict CORS to known origins.
- Set up rate limiting on
/api/generateand/api/providers/keys. - Enable the policy engine's strict mode for any regulated domain.
- Run the audit log to a tamper-evident sink (append-only storage, SIEM).
- Disable
localmode if untrusted users have accounts.
- Security reports: GitHub private vulnerability reporting
- General questions: GitHub Discussions
- Non-security bugs: GitHub Issues