Agent-ready workflow · ai

AI Agent and Tool Trust

Trace untrusted content through prompts, retrieval, memory, tools, approvals, and output sinks to a measurable controlled effect.

aillmmcptoolsprompt-injectionagents
Stable IDworkflow:ai-agent-tool-trustSource record dateHuman reviewNot recordedProvenanceCurated workflow graph

Enter with artifacts, leave with evidence.

Use when
  • RAG
  • tool call
  • MCP server
  • agent memory
  • browser action
  • external content
Prerequisites
  • controlled content
  • test account
  • tool privilege inventory
  • effect rollback
MCP retrieval

Agents can search this workflow, retrieve the complete graph, or request one stage through the read-only Vulns.co MCP connector.

01

Map trust and privileges

Record system instructions, user input, retrieved content, memory, tool schemas, approvals, secrets, and output consumers.

Inputs
  • approved AI feature
Outputs
  • trust-boundary map
  • tool privilege table
Evidence gate
Every tool has inputs, allowed effects, approval state, and data access noted.
Negative control
Plain model influence is separated from tool or sink effects.
Stop condition
No uncontrolled emails, messages, purchases, or external changes.
02

Plant a benign instruction canary

Use controlled content to test whether untrusted instructions cross a named boundary.

Inputs
  • controlled document or record
Outputs
  • instruction-following evidence
Evidence gate
The canary identifies its source and does not request secrets or privileged actions.
Negative control
A matched document without the instruction produces normal output.
Stop condition
Prompt influence alone is not reported as security impact.
03

Test one controlled effect

Ask whether the influenced model can cause a reversible tool call, unsafe render, URL fetch, or memory change.

Inputs
  • supported injection path
  • controlled effect target
Outputs
  • tool or sink evidence
Evidence gate
An unauthorized effect is observed and attributable to the controlled input.
Negative control
Approval, policy, or schema control blocks the same action when intact.
Stop condition
No secret disclosure or actions affecting third parties.
04

Bound the chain

Separate instruction following, tool selection, argument control, authorization, effect, persistence, and data access.

Inputs
  • effect evidence
Outputs
  • prerequisite graph
  • minimal report
Evidence gate
Each chain edge is observed or clearly labeled untested.
Negative control
Remove the untrusted content and repeat the controlled action.
Stop condition
Clean up memory and controlled state.

Continue with the right depth.

References