Skip to content

Security: GenAI-Security-Project/agent-control-standard

SECURITY.md

Security Policy

ACS is a specification. Most of what lives here is prose and JSON Schema, so the security surface is narrower than a typical software project. Report anything you find anyway. A flaw in the specification propagates into every implementation that follows it.

Report a vulnerability

Do not open a public issue.

Use GitHub's private vulnerability reporting. Reports land with the maintainers and stay private until we publish an advisory together.

Include what you have:

  • What you found and where, with a file path or a link to the specific line
  • How to reproduce it, or the reasoning chain if the flaw is in the specification rather than in code
  • What an attacker gains
  • Any suggested fix

Partial reports are welcome. We would rather triage something incomplete than never hear about it.

What is in scope

In scope Out of scope
Flaws in the ACS specification that lead implementers into insecure designs The documentation site at agentcontrolstandard.ai, which is built from a separate repository
Errors in the JSON Schemas under specification/ Findings against third-party agent frameworks that happen to implement ACS
The GitHub Actions workflows in .github/workflows/ Automated scanner output with no demonstrated impact
Hook or event definitions that leak sensitive data by design Missing security headers on sites we do not operate
Supply-chain issues in this repository's dependencies Social engineering of maintainers or contributors

A specification flaw counts. If a hook definition forces implementers to log secrets, or an event schema makes an authorization bypass easy to write, that is a finding even though no code here executes.

Response commitments

Stage Target
Acknowledge your report 3 business days
Initial triage and severity assessment 10 business days
Fix or documented mitigation for high and critical findings 90 days from triage
Public advisory Coordinated with you, at or before day 90

We use CVSS v3.1 for severity. Specification flaws get scored against a reference implementation, since the specification itself has no runtime.

Coordinated disclosure

We publish an advisory once a fix ships, or at 90 days from triage, whichever comes first. If a fix needs longer, we tell you why and agree a new date rather than letting the clock run out quietly.

Tell us how you want to be credited, including if you prefer not to be. We credit reporters in the advisory by default.

If you believe a finding is being actively exploited, say so in the report. That moves it to the front of the queue and shortens every window above.

Safe harbor

We will not pursue or support legal action against research conducted under this policy, provided you:

  • Report through the private channel above and give us a chance to fix the issue before disclosing it
  • Avoid privacy violations, data destruction, and interruption of any service
  • Access only the minimum data needed to demonstrate the finding, and delete it once you have reported
  • Do not exploit a finding beyond what proving it requires

Work in good faith under these terms and we treat your research as authorized. If a third party brings action against you for research that followed this policy, we will make that authorization clear.

Signing and provenance

Contributors sign off commits under the Developer Certificate of Origin. See CONTRIBUTING.md.

Report suspected compromise of a maintainer account or a release artifact through the private channel above, marked urgent.

There aren't any published security advisories