Prompt Injection Testing

Who it's for

Who this is for

Teams shipping agents that process untrusted content.
Security teams validating an agent before launch.
Engineering teams that want a repeatable injection eval suite.

Problems

Problems solved

Agents that follow instructions embedded in untrusted input.
Single-filter defenses that miss output- or argument-level attacks.
No repeatable eval suite tied to releases.
Unknown exposure to exfiltration through tool arguments.

Scope

What I review / build

Assemble an attack pack tailored to the agent's tools and inputs.
Run injection, tool-abuse, and exfiltration scenarios.
Test input, output, and tool-argument boundaries separately.
Measure detection and blocking per category.
Hand over the eval suite so the team can re-run it.

Deliverables

Tailored prompt-injection attack pack.
Test results with reproductions and severities.
Category-level detection/blocking metrics.
A re-runnable eval suite for CI.
Recommended layered-defense controls.

Process

Sample workflow

Attack design

Build scenarios specific to this agent — its tools, its untrusted inputs, its data.

Execution

Run the pack against a non-production instance across all three boundaries.

Scoring

Score by category; flag any successful unauthorized action or leak.

Handover

Deliver results plus a re-runnable suite the team owns.

Controls

Security controls included

Input-boundary detectionOutput-boundary filtersTool-argument validationBlock-or-review escalationContinuous red-team eval in CI

Typical effort

1 weekfocused effort

Scoped around availability. Async-friendly.

How to start

Get in touchno sales call

This is work I do and write about in the open — reach out to compare notes or ask a question.

FAQ

Frequently asked questions.

Is one classifier enough to stop injection?

No. The testing assumes defense in depth and probes input, output, and tool-argument boundaries separately, because no single filter catches every case.

Do I keep the test suite?

Yes. The deliverable includes a re-runnable eval suite tied to releases so your team can keep testing as the agent changes.

Is this authorized testing only?

Always. Testing runs against non-production instances with written scope. I do not perform unauthorized testing.

Want to talk it through?

Building or securing something similar? Tell me about your agent, stack, or SOC workflow — I'm happy to compare notes.

Get in touch All Services