Coordinated disclosure policy
agent-audit-kit publishes the MCP Security Index, an automated leaderboard grading public MCP servers. Before a finding is ever visible on that leaderboard, we follow this policy.
Reporting to maintainers
When our weekly crawl discovers a previously-unseen finding, we:
- Open a private security advisory (or private issue) on the affected repository as soon as we reasonably can after the scan that discovered it — best-effort, prioritised by severity, with no fixed clock (this is a solo-maintained project).
- Include: the rule ID (e.g.
AAK-MCP-011), the file + line pointer, the remediation text the scanner carries, the CVSS-estimated severity, and a link to this policy. - If no private-issue channel is available, we email the addresses
listed in
SECURITY.md/security@<domain>/ the last-committing author address — in that order. - Record the notice in
benchmarks/disclosure_ledger.json: the date and the channel, nothing else. That file is public, so it never holds rule IDs, email addresses or advisory links. The index reads its 90-day clock from this record and from nothing else.
Disclosure timeline
- Day 0: the private notice is sent and recorded in the ledger. A server whose maintainer has not been notified has no Day 0: its rule-level detail stays withheld however long it has been in the index.
- Day 30: reminder if no response or fix yet.
- Day 60: second reminder.
- Day 90: the server's rule IDs and per-rule counts are added to its public card on the MCP Security Index, with the first weekly snapshot on or after day 90 that includes the server. File and line locations are not published. The maintainer is notified again 24 hours before publication.
- Maintainer fix earlier: if the maintainer ships a fix before
day 90, we record the date as
fixed_atand the detail is published with the next weekly snapshot after it, with a thank-you credit.
Reminders, the 24-hour heads-up and the credit are sent by hand. The index automates only the dates: it withholds detail until the ledger says the 90 days have passed or a fix has landed.
What we publish during embargo
Until then, a server's grade can still shift (e.g. from B to
C), but its public card and its row in data/index.json show only
the aggregate severity counts, not the rule IDs. Any research-grade
detail is held until embargo expiry.
What we do not do
- We do not publish proof-of-concept exploits.
- We do not publish findings against projects under active coordinated disclosure with another party (we honor the earliest embargo).
- We do not accept bug bounties.
Contact
Security reports about agent-audit-kit itself: open a private advisory at https://github.com/sattyamjjain/agent-audit-kit/security/advisories. This is the only supported channel. It creates a timestamped, triaged record, and the 90-day clock for a report about agent-audit-kit runs from that record.
Scope
This policy applies only to detections produced by agent-audit-kit's automated scanners and to the MCP Security Index published from those scans. Hand-authored research we happen to come across is reported via the affected project's own policy and has its own timeline.
Changes
| Date | Change |
|---|---|
| 2026-04-18 | Initial version (v0.3.0 launch). 90-day embargo. |
| 2026-10-10 | The clock starts at the recorded private notice (benchmarks/disclosure_ledger.json), not at first scan. Until 2026-10-10 the index restarted it every weekly run, so nothing ever reached day 90, and data/index.json carried the rule IDs of withheld servers. Servers not yet notified now stay withheld, data/index.json carries no rule IDs while a server is withheld, and the doc says which steps are manual. |