Prompt injection
Malicious instructions introduced through user input or content an AI system processes, with the aim of redirecting its behaviour.
Agent and tool call
An agent combines model output with workflow logic and capabilities. A tool call requests an operation; the destination log helps establish whether that operation actually occurred.
RAG and persistent memory
Retrieval-augmented generation supplies source material to a model. Persistent memory stores information across runs. Each introduces source, permission and retention questions during an investigation.
Containment and recovery
Containment limits ongoing impact. Recovery restores an approved service state and validates the conditions required to operate it. Neither term implies that every historical action can be reconstructed.
Attempt versus confirmed effect
A tool request shows what was attempted. A destination record can show whether the operation completed or changed state. Those observations should be kept separate in an incident report.
Recovery acceptance gate
The agreed checks and approval required before a workload returns to operation. The gate should include the implicated failure condition, legitimate use, residual risks and an accountable service owner.
MCP
Model Context Protocol connects AI applications with tools and context sources. The investigation depends on the deployed client, server, authorisation and downstream permissions. See MCP investigation scope.
