AI agent forensics glossary

Which questions belong in this field?

Define the action being investigated and the environment in which it could occur. Writing a proposed command is different from submitting it to a tool; a submitted request is different from a completed operation. These stages can leave different evidence.

Record permission boundaries and approval context. A runtime may require approval for one operation but permit another automatically. A visible approval event supports a specific authorization claim only when its scope and relationship to the action are understood.

Evidence sources and examination context

Connect tasks, requests, tool responses, and target-system records through identifiers and supported timing. Preserve the available payloads and outcomes while respecting the authorized collection scope.

The Model Context Protocol Tools specification describes tool calls and results for a particular protocol revision. It does not require a persistent forensic audit log. Examine what the implementation actually retained rather than treating the protocol as a guarantee of recoverable history. Model Context Protocol: Tools, specification 2025-11-25.

For retries and delegated tasks, distinguish logical intent from repeated executions. Compare receipts with the state of the destination where possible. A success response can be informative without proving that the intended final state persisted.

A worked evidence example

Hypothetical example: an agent requests that a file be uploaded, receives a timeout, and retries. One later success message appears in the conversation. The destination may contain one upload, two uploads, or neither if the state later changed.

Request identifiers, target records, and file comparisons help distinguish those accounts. The examination should not count tool-call messages as completed uploads or assume that a retry is harmless merely because the interface shows one result.

What this evidence can and cannot establish

Agent records can support findings about recorded requests, permissions, reported outcomes, and task relationships. Target-system evidence can corroborate particular effects.

An execution trace is not necessarily complete. Missing logs do not establish that no action happened; a logged action does not establish a person’s intent. Attribution to an account, an agent configuration, and an individual are separate conclusions.

Essential terminology

• Agent — An AI-enabled system that can pursue a task through actions, such as using tools, under a particular runtime and permission context.
• Agent forensics — Examination of AI-driven task execution, tool use, state, delegation, and effects on local or external systems.
• Tool call — A recorded request to a tool with particular arguments and execution context. A request alone does not establish the external effect.
• Delegation — Assignment of work from an initiating agent or task to another agent or process, with a traceable relationship where records permit.
• Approval event — A record of a decision authorizing a specified action or change.
• Execution trace — Records of steps performed during a program or agent run.
• Idempotency — A property in which repeating an operation has the same relevant effect as applying it once.
• Corroboration — Support for a proposition from independent evidence, with independence assessed for the particular claim.

All 32 terms in this evidence field

• Agent
• Agent forensics
• Agent task
• API audit event
• Approval event
• Asynchronous operation
• Command execution
• Correlation
• Corroboration
• Delegation
• Delegation chain
• Execution trace
• Exfiltration
• Exit status
• Function calling
• Idempotency
• Job queue
• Least privilege
• MCP
• Network request
• Parent process
• Permission boundary
• Process identifier
• Race condition
• Retry
• Sandbox
• Standard error
• Standard output
• Task graph
• Tool call
• Tool invocation
• Tool result

Continue with an examination reference

• agent forensics
• corroboration
• reconstruction

Citation and editorial context

This is an editorial reference. Source checks are separate from independent expert review; no independent reviewer is recorded.

Examples are hypothetical. See the editorial policy, methodology, and correction process.

Shelley, Jesse (ed.). “AI agent forensics glossary.” Synthetic Forensics, 1.1.0-glossary-library-1, 2026-10-09. https://syntheticforensics.org/glossary/agents/ Accessed: add the date you consulted this page.

The complete vocabulary

Browse the full glossary of 300 terms or download the machine-readable vocabulary.

Field assignment is the project’s organizational taxonomy; related evidence can span several fields.