Security policy

oRPC Agent is a security-focused framework with v0.1 implemented in-repo but not yet published: reports may target either the design or the code, and we want both.

Reporting

Do not open public issues for exploitable problems.

• Preferred: GitHub private vulnerability reporting ("Report a vulnerability") on this repository.
• We aim to acknowledge within 72 hours and give an assessment within 14 days. Coordinated disclosure: we'll agree on a timeline with you; default publication is upon fix release (or 90 days, whichever is sooner).
• No bug bounty exists; credit is given in advisories unless you prefer otherwise.

In scope: violations of the documented security invariants (SI-1…SI-12), flaws in the approval integrity design, concealment gaps ([SI-8]), error-information leaks (SI-9), unsound claims in the docs, and — now that code exists — anything where the implementation lets an agent surface do what the documentation says it cannot.

Out of scope: vulnerabilities in applications built with the framework (report to those projects), in oRPC itself, in model providers, and prompt-injection occurrence (the framework bounds impact; it does not prevent injection — documented).

What the framework does and does not claim

Read before reporting, to calibrate expectations:

• It provides controls for exposure, execution-time authorization, approvals, redaction, and audit; it requires application-level authentication and authorization to exist.
• It does not claim to solve prompt injection, provide exactly-once execution, or be "production-safe by default".
• Supported-version policy (post-release): latest minor receives fixes; the previous minor receives critical fixes for 6 months.

Design-phase security review

Structured review is actively wanted: the threat model (T1–T15) and the execution pipeline are the highest-value review targets. Non-exploitable design critique is welcome as a public issue labeled security-design.