About the project
What DiggingBeagle does
An agent can cross a trust boundary through a tool, a shared service, or a copied instruction. The route matters more than the headline.
The project
DiggingBeagle is a non profit research project that accumulates documented AI security incidents, agent failures, vulnerabilities and AI-assisted operations. It publishes case reconstructions, dated news, articles that compare records, and updates when an account changes.
The work is meant to be inspected. A case states what happened or was demonstrated, identifies the mechanism and links individual claims to sources. The page keeps unresolved questions beside the account instead of hiding them in a summary.
What enters the archive
A real incident, a measured experiment, a simulation and a reported allegation can all teach something different. They remain labeled as different evidence types. We do not turn a test result into an incident or treat a named person or company as responsible merely because a record links to them.
Topics connect records only when the published material makes that relationship explicit. A topic count is a map of those links, not a score for severity or importance.
How publication works
Research proposals can be assisted by software, but a human operator reviews canonical changes and separately approves material for publication. The public site is an activated, immutable snapshot. Later research appears through a new reviewed publication, and current restrictions can remove material from the current site.
When an account changes
New evidence can narrow a claim, contradict an earlier report or establish a response that was not previously known. We keep dated updates and corrections distinct from news so a reader can trace the change. The methodology describes how claims, sources and uncertainty are handled.