Prompt injection and AI trust-boundary abuse
V10 · reusable vector · version 1.0MED · proposed test impact
Untrusted content attempts to become an instruction or an authority for a model-connected action. The policy must be enforced outside the model.
Attacker tools and payload types
Hostile natural-language instructions in a document/message. The assessment uses the supplied inert fixture and synthetic client records.
The entries below are tools that can support assessment or emulate aspects of this vector in an authorised fixture. They are not claims that these exact products were used in every linked incident. Native platform availability and Armory installation are different from tested capability.
- AgentShield · native service; handoff is not automatic execution
- Decepticon · native service; handoff is not automatic execution
- PentestGPT · native service; handoff is not automatic execution
Repository and operating guides
ai-fixture.txt · RES-023 · published synthetic fixture in the assessment collection linked from the Repository.
Prompt Injection/README.md · existing Repository reference; choose only relevant approved checks.
Read the selected tool’s guide for installation and operation rather than recreating it here. Pin the deployed version and reviewed configuration in the task record. Some local guides are autogenerated inventory notes and contain stale commands or broad prerequisite lists; they require validation before use. The official manuals linked above were checked during this review.
Reusable controlled test
Run matched hostile and benign cases against a sandbox with non-destructive tools. Observe responses, attempted tool calls and the independent allow/deny result.
- Required permission
- Exact assistant, dataset, model budget, test tools and permitted actions; retrieved text never grants authority.
- Parent supplies
- Scenario/risk ID; business question; exact targets; named permitted people; channel; current policy revision; tool/resource versions; limits; expected result and reviewer.
- Output
- One result per approved case, with run and target IDs, permission decision, evidence, actual outcome and coverage limits.
- Measure
- Unauthorised actions blocked / hostile cases; useful benign task completion; policy-change and interruption outcomes. No client result has been collected here.
- Failure handling
- Blocked, not authorised, failed, inconclusive and not tested remain visible. A refused route cannot be replaced by another agent, channel or target without fresh applicable approval.
What the vector returns to the scenario
| Result | What it means |
|---|---|
| Control held | The evidenced attempted action was prevented under the tested conditions. This is not a universal security guarantee. |
| Control failed | Describe the exact observed boundary failure and its business consequence, without inflating it into a wider compromise. |
| Unknown / not tested | Record the missing capability, evidence, permission or coverage. Do not invent a score. |
Mitigation recommendations
Reduce likelihood: Keep trust boundaries explicit and enforce data/tool access independently.
Reduce impact: Limit accessible data and tool effects and retain reviewable evidence of consequential actions.
Define how a later retest would verify the recommendation; implementation assignments, scheduling and pricing are Part 2.