Skip to content

Security: swarmproof/agent-postmortems

Security

SECURITY.md

Security policy

agent-postmortems is a public corpus of AI-agent security incidents. Almost everything in it is about vulnerabilities, so this page is specific about what counts as a security issue in this repository and how to report one.

Scope

Report these here:

  • The tooling and workflows. Anything under scripts/, schema/, tests/, .github/workflows/, or the site build that could be abused: for example, a validator or export script that executes untrusted input from a record or a source archive, or a workflow that leaks a secret or accepts unreviewed content into main.
  • A record that carries a runnable exploit payload. Records describe a failure class and a trigger shape in natural language. If an incident file, a source archive, or an export (export/*.json) contains a working payload, prompt-injection string, or exploit chain that could be copied and run, that is a defect in the corpus.
  • A record that carries personal data beyond its published sources. Records never identify harmed individuals beyond what the cited public sources already published. Anything that goes further is a defect, whether or not the information is accurate.

Not in scope here:

  • The incidents themselves. A vulnerability in a vendor's agent, framework, or model belongs with that vendor. This repository only documents what public sources already report.
  • Corrections to a record's facts. Those follow DISPUTES.md.
  • Dead links, schema errors, or taxonomy questions. Those are ordinary bugs: open a normal issue.

How to report

GitHub private vulnerability reporting is not currently enabled on this repository, so there is no "Report a vulnerability" button on the Security tab.

Until it is:

  1. Open a GitHub issue titled Security contact request and say only that you have a private report. Do not put details, payloads, or the affected file in the issue.
  2. A maintainer will open a private security advisory draft and add you as a collaborator on it. Details go there.

If private vulnerability reporting is enabled later, the "Report a vulnerability" button on the Security tab becomes the preferred route and this page will be updated.

There is no security email address for this project. Any address claiming to be one is not ours.

What we do not accept

  • Runnable exploit payloads in records or exports. The curator standard and the export contract both require natural-language trigger shapes only. A pull request that adds a working payload will be asked to remove it, regardless of how well-sourced the incident is.
  • Proof-of-concept code against third-party systems. Link to the public writeup; do not reproduce the exploit here.
  • Reports about vendors' systems addressed to us. We cannot fix them and we will not relay them.

Response

This project is maintained by volunteers. There is no SLA.

  • A Security contact request issue is normally acknowledged within a week.
  • A confirmed payload or personal-data defect in a record is treated as the highest priority the maintainers have and is fixed by editing the record, never by deleting it (see ADR-0008).
  • Tooling fixes ship as ordinary pull requests once a fix is ready. If a workflow is the problem, it may be disabled first and fixed later.

We will credit reporters in the fixing pull request unless asked not to.

There aren't any published security advisories