oya is pre-1.0 and ships from main. Security fixes land on the latest published oyadotai / oyadotai-server versions on npm. Please always test against the newest release before reporting.

Do not open a public issue for security problems.

Report privately through either channel:

• GitHub - open a private advisory (Security → Advisories → Report a vulnerability). Preferred.
• Email - mk@oya.ai with steps to reproduce, affected versions, and impact.

We aim to acknowledge reports within 3 business days and to ship a fix or a mitigation plan within 90 days, coordinating disclosure timing with you.

oya's core security claim is the projection lattice: values marked OPAQUE must never reach the planner-facing view. We are especially interested in:

• Any input that causes an OPAQUE value (or a sentinel planted in one) to appear in the model-facing context - this is the headline invariant (packages/core/test/executor.test.ts) and must never be weakened.
• Static-checker bypasses (packages/core/src/checker.ts) that let a plan execute with a disclosure it didn't declare.
• Sandbox escapes in WorkerSandbox / InProcessSandbox.
• Prompt-injection paths that reach the planner through tool output.

Out of scope: issues in your own tool execute implementations, third-party model providers, and the hosted platform at oya.ai (report those to mk@oya.ai directly).

Thank you for helping keep oya and its users safe.