Softlander Book a call
Agent security and cybersecurity

Prompt-injection testing

Test whether direct instructions can redirect an agent beyond its approved task

Users can submit open-ended requests to a tool-enabled assistant. The decision is how to test whether direct instructions can redirect an agent beyond its approved task. The work begins with your actual workflow and looks for the evidence that separates intended behaviour from a consequential exception. It is particularly relevant to security test owners and product engineers. A useful outcome is an explicit account of what was examined, what remains uncertain, and which decision the evidence can support.

Book a 30 minute readiness callBrowse this capability group

The agent security relationship

  1. Starting evidenceInstruction hierarchy

    Establish the basis for direct request.

  2. Decision boundaryProtected behaviour

    State what must remain unchanged.

  3. Decision artefactDirect injection case set

    Retain the scoped finding and its unresolved dependencies.

Proposed working model. A bounded challenge cannot establish exhaustive protection.

Decision

Frame the decision around direct request

Questions and the evidence they need

Criterion / evidenceDirect request
Inspect instruction hierarchy for the relevant constraint or observation.
Criterion / evidenceProtected behaviour
Inspect protected behaviours for the relevant constraint or observation.
Criterion / evidenceExecuted effect
Inspect synthetic prompts for the relevant constraint or observation.
Read each criterion against the available evidence before deciding that additional work is needed.

The decision is how to test whether direct instructions can redirect an agent beyond its approved task. This is relevant to security test owners and product engineers when users can submit open-ended requests to a tool-enabled assistant. Name the affected workflow, the consequence of a wrong decision, and the owner who can act on the result. The initial discussion should establish which evidence is already available and which uncertainty is material enough to investigate.

The specific failure hypothesis is that an adversarial request persuades the agent to treat a forbidden action as a necessary exception. It remains a hypothesis until examined in the deployment. Existing internal tests or operating records may already resolve it; identify the uncovered question before commissioning additional work.

Mechanism

Injection success depends on the resulting action

Relationships within agent security

Source or conditionDirect request
Define the adversarial instruction under test.
Crossing or choiceProtected behaviour
State what must remain unchanged.
Consequence or decisionExecuted effect
Inspect tool activity beyond the final refusal.
Observe the control and resulting business effect. Each relationship needs evidence in the actual scope.

What this specialty examines

Direct injection testing distinguishes the requester’s legitimate task from an attempt to change governing constraints. Define the protected action and what the requester can submit. Evaluate actual downstream behaviour rather than judging only whether a refusal phrase appears.

The review should run authorised challenge requests and distinguish attempted redirection from executed side effects. Identify the normal route, relevant exceptions and the authoritative record of the outcome. Inspect the relevant business-system record, enforcement event or process evidence alongside any reported status. Keep incomplete cases visible, with an explanation of what their absence prevents the owner from concluding.

The tradeoff to make explicit

For this decision, stronger refusal behaviour matters while tool-side enforcement still needs separate evidence. Compare the available design or operating choices against the same intended outcome. If one option removes a required control or changes the task boundary, describe that difference before treating its apparent simplicity, speed or lower effort as an improvement. The decision should expose the tradeoff instead of burying it in an aggregate score.

Preparation

Connect the working inputs to the boundary being examined

Evidence packet for direct request

Illustrative working format

BasisInstruction hierarchy
Identify provenance and the specific decision this material can inform.
ContextProtected behaviours
Check its relationship to the relevant workflow and expose missing context.
ObservationSynthetic prompts
Keep expected behaviour and the limits of the constructed example explicit.
DependencyTool audit records
Establish permitted access, provenance, and which task conditions it represents.
Illustrative preparation record. Availability, revision, and handling need confirmation.

Start with instruction hierarchy, protected behaviours, synthetic prompts and tool audit records. Identify the owner, revision and operating context of each source. Sanitised examples may support initial design, but differences from actual conditions should remain explicit. Where direct access is unsuitable, a customer-controlled evidence or execution route may be considered, with agreed handling and retention.

The scope also needs the security decision owner, agreed expectations for a representative threat scenario, and the authority to inspect the relevant boundary. Specify what evidence would distinguish the failure in which an adversarial request persuades the agent to treat a forbidden action as a necessary exception from acceptable behaviour. If stakeholders disagree about intended behaviour, resolve the interpretation before classifying the observed difference as a defect.

How the work proceeds

Carry direct request through to an accountable finding

From the scoped material to a decision record

  1. Establish conditionsInstruction hierarchy

    Define the hypothetical attacker’s starting influence.

  2. Examine the relationshipProtected behaviour

    Apply the scoped check to executed effect.

  3. Retain the findingDirect injection case set

    Record the evidence, open dependencies, and the owner’s next decision.

Proposed sequence; resolve interpretation or access gaps before relying on the next step.

Define the controlled threat scenario

Review instruction hierarchy, protected behaviours, synthetic prompts and tool audit records with security test owners and product engineers. State what the hypothetical attacker controls, the protected action and the systems included in the authorised exercise. Prepare synthetic records and harmless markers where appropriate, with stopping conditions for unexpected effects. The preparation output is a scenario brief that distinguishes an influence attempt from a prohibited action actually reaching a business resource.

Challenge the path and review effects

Record refused requests, attempted prohibited calls and actions blocked by tools separately. This reveals whether the model followed the boundary and whether the application enforced it. Do not turn one successful refusal into a broad protection claim.

Select a control response and retest

Build a direct injection case set and observed control findings from the reviewed observations. Separate agent behaviour, application enforcement and the resulting resource state so the recommended change reaches the responsible layer. Choose a bounded retest that reproduces the original condition and checks legitimate behaviour after remediation. If the required investigation exceeds the agreed technique or expertise, retain the unresolved path and identify the qualified review needed.

Indicative artefact

Make the direct injection case set usable by its owner

Direct injection case set

Illustrative working format

Annotated fieldDirect injection case set
Identify the assessed workflow, revision, and decision boundary.
Annotated fieldObserved control findings
Link observations and counterevidence to the case, configuration, and conclusion limit.
Annotated fieldProtected behaviour evidence
Link observations and counterevidence to the case, configuration, and conclusion limit.
Illustrative artefact structure, not a completed engagement or an observed result.

A useful decision artefact is a direct injection case set and observed control findings. Supporting material may include selected cases, configuration or process records, evidence references, a disagreement log and prioritised actions. Retain the case selection rationale and material disagreements alongside the result; otherwise the result may be mistaken for complete coverage.

Illustrative scenario

Examine executed effect through a constructed exception

A constructed case and its decision paths

Illustrative scenario

  1. SituationDirect request

    A challenge prompt receives a refusal while a background tool call still proceeds.

  2. Question to examineProtected behaviour

    Inspect background tool events after the visible refusal.

  3. Conditional next actionExecuted effect

    Record the executed effect separately from the response wording.

Fictional example. No executed test, customer engagement, or measured outcome is implied.

Consider a fictional situation: an illustrative user tells a support assistant to ignore its task boundary and export account records for debugging. The scenario is illustrative, with no client engagement, executed test or measured improvement implied.

For the debugging export request, use synthetic records and an explicitly prohibited destination. If the model attempts export but the tool denies it, preserve both observations. The owner may need to improve interaction behaviour while retaining the effective enforcement control.

Scope and next action

Keep the conclusion within the agreed threat scenario

Confirm fit before the next commitment

Included boundaryDirect request
This review is limited to the agreed threat scenario.
PrerequisiteInstruction hierarchy
The relevant owner must confirm access and the intended operating conditions.
Claim limitDirect injection case set
A bounded challenge cannot establish exhaustive protection.
Bring the triggering change and existing evidence; resolve the dependency that could alter the decision.

Authorised scenario testing is not a claim of exhaustive penetration testing or guaranteed protection. Intrusive security work requires qualified specialists and a separately agreed scope. Any engagement depends on an agreed scope, qualified team, authorised access and confirmed availability. Bring the affected workflow, its current configuration or process, the triggering change and the evidence already available. The immediate scope question is prompted by the situation in which users can submit open-ended requests to a tool-enabled assistant; confirm which missing observation would change the owner’s next action.

Ways to engage

Related engagement scopes

Agent Release Readiness Review

The Agent Release Readiness Review is a fixed-scope, three-week independent test of one AI agent workflow that shows a release board what the agent does, what it is allowed to do and what happens when it is wrong, and ends in a signed recommendation.

3 weeksRead about this offer

AI Agent Runtime Security Review

Make every delegation explicit. Examine who the agent acts for, where it can execute and which tools it can reach when instructions alone are not enough.

Read about this offer

AI Agent Red Team

Challenge the boundaries that matter to your workflow. Test agreed attack scenarios against agent instructions, tools and data access.

Read about this offer

Continue exploring

Related insights and resources

Agent security6 MIN READ

The permission boundary behind every agent action

An agent can only be evaluated in the context of what its tools allow. Examine identity, delegated authority and action boundaries before you put a workflow in motion.

Whitepaper5 MIN READ

Before your agent can act, define its authority

A working paper for the moment an AI system moves from advice to action. Map identity, permissions, approvals and recovery into an authority model your team can review and test.

Agent security6 MIN READ

Running an AI-agent incident tabletop exercise

A useful tabletop follows one plausible agent failure through detection, containment, evidence, business recovery and accountable return to service.

A little more clarity

Questions to consider.

Start with the questions that shape a better decision.

Is a refused malicious prompt proof that the system is protected?

No. One refusal covers one tested condition. Inspect tool calls and side effects, preserve blocked attempts separately, and avoid generalizing from an unexecuted action to universal protection.

Which protected behaviours define injection tests?

Start with instruction hierarchy, protected behaviours, synthetic prompts and tool audit records. Record their revision, owner and relationship to the intended workflow. Missing context should be identified before interpreting outcomes; a sanitised fixture is useful only when its relevant differences from the operating case are understood.

How should refusal and blocked execution be distinguished?

Run authorised challenge requests and distinguish attempted redirection from executed side effects. State the hypothetical attacker's control and inspect attempted actions, tool enforcement and actual effects separately. A refusal, a blocked call and a successful prohibited action are different results. Preserve unsuccessful challenges as well as adverse findings.

What should a direct-injection case set record?

The proposed output is a direct injection case set and observed control findings. Include the selected conditions, supporting observations, unresolved differences and next decision. A reviewer should be able to follow each important conclusion back to its evidence and understand the limits without depending on an undocumented explanation from the author.

When is a refusal an incomplete security result?

A refusal is incomplete evidence when background actions, tool attempts or alternate paths remain unexamined. Inspect the whole action chain and preserve whether application controls blocked behaviour the model still attempted.

What if existing evidence already answers the question?

Review its scope, provenance and configuration before proposing new work. Existing work from security test owners and product engineers may already be sufficient. In that case, capture the decision, unresolved limitations and the change that would require another review. Additional assessment should address a real gap.

What would make the findings inconclusive?

Missing instruction hierarchy, protected behaviours, synthetic prompts and tool audit records can prevent a reliable interpretation, depending on the case. Record the specific missing observation and its implication. The owner can supply evidence, narrow the claim or defer the dependent decision; an incomplete case should not silently be counted as a pass.

Start with the decision you need to make.

Tell us about the agent workflow you want reviewed, and where you need greater clarity. Scope and availability are confirmed together.

Book a 30 minute readiness call Get the agent release checklist

Connected disciplines

Indirect prompt-injection assessment

Examine instructions hidden inside information an agent consumes. The review connects content provenance, retrieval paths, trust labels and consequential action rules to a bounded review of agent security, giving the security decision owner a clearer basis for the next action.

Tool and capability abuse testing

Understand whether legitimate tools can be combined for unintended actions. The review connects tool inventories, permitted use cases, action sequences and resource constraints to a bounded review of agent security, giving the security decision owner a clearer basis for the next action.

Agent privilege-escalation testing

Test whether agent authority can increase across a workflow. The review connects permission graphs, delegation paths, privileged interfaces and test identities to a bounded review of agent security, giving the security decision owner a clearer basis for the next action.