Security fixes are applied to the latest version on the default branch.

Please report vulnerabilities privately and do not open public issues for active security problems.

Contact:

• anton@kay.ai

When possible, include:

• Affected component and version/commit
• Reproduction steps or proof of concept
• Impact assessment
• Suggested mitigation (if available)

We will acknowledge receipt and triage as quickly as possible.

After a fix is available, we may publish a summary describing impact and remediation guidance.

Source documents are included in model prompts and must be treated as untrusted input. Prompt-injection text can influence both count estimates and extracted values. Count anomaly fallbacks reduce obvious failure modes but cannot detect all plausible under-counts or fabricated output.

For adversarial inputs, use task-specific validation and an independent coverage check, keep model credentials scoped, leave tracing disabled unless appropriate, and do not execute or directly trust extracted content.