Only the latest CalVer release receives security patches.

Do not open a public issue for security vulnerabilities.

Email javier@jmarin.info with:

• A description of the vulnerability
• Steps to reproduce
• Affected versions
• Any potential impact assessment

• Acknowledgment: within 48 hours of your report.
• Initial assessment: within 5 business days. We will confirm whether the report is accepted, request additional information, or explain why it does not qualify.
• Fix timeline: critical vulnerabilities within 7 days, high severity within 14 days, moderate within 30 days.
• Disclosure: we will coordinate disclosure timing with you. We follow responsible disclosure practices and will credit reporters unless they prefer anonymity.

The following are in scope:

• Code execution vulnerabilities in groundlens core, CLI, providers, or integrations
• Dependency vulnerabilities in direct dependencies (numpy, sentence-transformers) that affect groundlens users
• Data leakage through calibration files, cached embeddings, or provider wrappers
• Deserialization attacks via crafted calibration JSON files or CSV inputs
• Path traversal through CLI file arguments or calibration file paths

The following are out of scope:

• Vulnerabilities in third-party LLM provider APIs (OpenAI, Anthropic, Google) -- report these to the respective providers
• Accuracy of hallucination detection scores (this is a research concern, not a security concern)
• Denial of service through large input texts (sentence-transformers has its own input limits)

• No secrets in code. groundlens does not store or transmit API keys. Provider wrappers read keys from environment variables or user-provided configuration.
• Dependency auditing. We run pip-audit in CI against known vulnerability databases.
• Type safety. mypy strict mode is enforced across the codebase.
• Input validation. All public functions validate inputs before processing.
• Minimal permissions. groundlens requires no network access for core functionality. Network access is only needed when loading sentence-transformer models (first use) or when using LLM provider wrappers.