The Kinetica Admin Agent is in pre-1.0 development. Security fixes target the latest 0.x release only. Once 1.0.0 ships, this policy will be updated to cover the active minor line.

Do not open a public GitHub issue for security vulnerabilities.

Email security@kinetica.com with:

• A description of the issue and the potential impact
• Steps to reproduce (minimal proof-of-concept is ideal)
• The affected version (output of admin-agent --version or the commit SHA)
• Your name or handle for acknowledgement, if you would like credit

We aim to acknowledge reports within 48 hours and to have a patch in testing within 14 days for critical issues.

As a backstop, you may also submit a private report through GitHub's Security Advisories flow.

We ask that you give us 90 days from initial report to publish a fix before public disclosure. For issues under active exploitation we will work with you on an accelerated timeline.

• Credential handling — Kinetica user/pass, Anthropic API key, OAuth tokens
• The mutation-tool approval gate and its bypass paths
• Prompt injection via untrusted data sources (database contents, table names, report text)
• Supply-chain concerns in this repository's build pipeline (dependencies, CI, release artifacts)
• Report-scrubbing completeness for saved diagnostic reports
• Downgrade attacks on URL resolution (the HTTPS → HTTP fallback path)

• Bugs in Kinetica itself — please file those with Kinetica support
• Bugs in the underlying Claude Agent SDK — file with Anthropic
• Issues requiring physical access to a user's machine or their pre-compromised Anthropic/Kinetica account
• Denial-of-service against the agent process itself (running a local CLI with hostile input you control)

This tool is an administrative diagnostic agent that runs with admin-level Kinetica credentials and executes approved mutations. It is not designed to be:

• Run as a multi-tenant service
• Exposed to untrusted users
• Run with unattended approval of mutation tools

Deploying it outside those intended bounds is not a supported use case, and vulnerabilities that only surface in such deployments may be closed as out-of-scope.