Fixed-scope assessment

AI Application Security Assessment

Your AI-enabled application may be able to expose sensitive data or take actions its user is not allowed to perform.

Understand and improve the security posture of your AI-enabled application.

Describe one AI feature

Does this apply to your product?

This matters if your chatbot, agent, MCP integration, or AI workflow can read sensitive data, call tools, or change application state.

We use AI surface as a short name for the part people or systems interact with: a chatbot, agent, MCP integration, harness, or workflow node.

Prompt injection is when untrusted instructions persuade an AI feature to ignore its intended rules or misuse the access the application gives it. It matters because the application around the model—not just the model—decides which identity, data, tools, and actions are available.

The decision this work supports

Know what the application actually permits

Before the assessment, you may know what the feature is intended to do without knowing what happens when its instructions, user permissions, service access, tools, and application controls disagree.

Afterward, you know which meaningful behaviours were allowed, attempted, denied, or completed; where the responsible control lives; what to change; and whether the correction worked.

The central question

What can someone persuade your AI surface to see, call, change, or disclose?

Evidence state explorer

A warning, a block, and a verified outcome are not the same thing.

Select each state to compare what appeared in the interface with what happened in the application.

What the model showed

A response that sounds cautious is shown to the user.

What the application did

The overpowered service identity still completes the unsafe email action.

What this proves

The application is vulnerable. The message did not enforce the boundary.

What the model showed

A detector raises a warning around the same attempt.

What the application did

The downstream action remains possible because nothing enforces the warning.

What this proves

Detection is evidence, not prevention. The consequential path is still open.

What the model showed

The interface reports that this pinned attempt was stopped.

What the application did

An application control denies the action before it reaches the email service.

What this proves

This path was blocked. It does not establish universal protection.

What the model showed

The same meaningful attempt is repeated after the correction.

What the application did

Authoritative readback shows that no unsafe email side effect exists.

What this proves

The narrow correction held for the pinned scenario and produced the intended effect.

What good testing establishes

From an interesting response to an application result

A model accepting an instruction is not the same as a harmful outcome. We declare the allowed and prohibited behaviour, follow the attempt through the application, inspect the authoritative downstream result, identify the owning control, and repeat the same path after correction.

The engagement includes assessment, a clear explanation of meaningful findings, practical remediation with the product team, and one bounded retest of the important corrected path.

A bounded worked example

The same synthetic HR scenario, observed four ways

The chatbot sent data Joe was not allowed to access.

Vulnerable baseline

The system raised a warning, but the data still left.

Detected, not blocked

An added control stopped this attempt before the email was sent.

Optional enforcement

We repeated the same test and verified that no email was sent.

Retest

In a synthetic HR chatbot demonstration, the same attempt is shown as a vulnerable service-identity baseline, detected without prevention, optionally blocked, and retested with authoritative readback showing no unsafe email side effect.

This one pinned source-and-target scenario does not prove that every prompt injection is blocked. Runtime is optional, and the example is not a claim of universal safety.

Read the evidence and limitations

What you receive

Evidence your product and engineering teams can use

  • A short leadership summary tied to business consequences
  • An application and trust-boundary map
  • Evidence-backed findings with safe reproduction steps
  • Prioritized remediation guidance for the product team
  • A remediation workshop
  • One bounded retest
  • Reusable test cases or evidence artifacts where safe and appropriate

Preconditions

What we need from you

An authorized target, a named owner, a suitable test environment or explicitly accepted production constraints, test identities, safe test data, relevant architecture context, and access to the team that can explain and change the application.

Exclusions

What this assessment is not

It excludes general infrastructure penetration testing, compliance certification, unrestricted production testing, destructive testing, and specialist reviews outside Viewyonder's competence.

Delivery

Scope, assess, explain, remediate, retest

  1. Scope: agree the target, roles, identities, tools, sensitive data or actions, evidence handling, exclusions, and success criteria.
  2. Assess: combine automated and manual testing for relevant instruction, authorization, tool-use, agency, disclosure, and control-bypass risks.
  3. Explain: reconstruct meaningful application paths and identify the identity, authority, capabilities, evidence, and control involved.
  4. Remediate: work with the builders on narrow corrections, improvements, accepted limitations, and specialist questions.
  5. Retest: repeat important paths and classify them as fixed, mitigated, accepted, or unresolved.

The assessment can combine manual testing, suitable third-party tools, purpose-built harnesses, Injectionator, or a combination. The method follows your AI surface and the evidence available from the application.

The assessment and remediation do not require Injectionator Runtime. Runtime is optional, and this service is not a product-adoption prerequisite.

Start with what you know

Describe one AI feature you are concerned about.

An ordinary-language description is enough to start. The remaining scoping questions are optional.

Start the assessment conversation