Every agent breach is a chain-of-custody failure.
AgentProv is the model that names the eight links, the data that shows which ones break, and the tools that check them.
Headline numbers
385
incidents in the catalog
2024 to 2026-10-27
325
incidents in 2026 so far
84% of the catalog
Pending
top layer share
Layer tagging in progress
The eight provenance layers
outermost boundary first- 1CredentialShould this run hold this key at all?
- 2InstructionWho actually said "do this"?
- 3ToolIs this tool what its description claims?
- 4Code / PackageWhere did this code come from?
- 5ConfigWho wrote the file that runs on open?
- 6Identity / RunWhich agent, run, or tenant is this?
- 7Build / ApprovalWas this approved as what it claims to be?
- 8Output / DestinationWhere is this data going, and who chose that?
The Model
Each layer, its question, and the attack patterns that live there.
The Data
Charts over 385 cataloged incidents, embeddable anywhere.
The Program
Tools that check specific layers, the catalog, and the annual report.