guest@apres:~/trainings/breaking-agents$ promptfoo redteam run

Training · Wednesday, February 24 · 6 hours

Breaking Agents: Hands-On AI Red Teaming and Evaluation Pipelines

Get a Gold Medal Pass $534.80 until Nov 30, plus Eventbrite feesAll trainings
Bryson Loughmiller

Instructor

Bryson Loughmiller
Founder, Petracore

Breaking Agents: Hands-On AI Red Teaming and Evaluation Pipelines

Learn to attack AI agents by hand, then make AI do the heavy lifting. You'll manually red team live agentic applications, direct an AI testing agent to recon them and build Promptfoo campaigns, and leave with an evaluation pipeline you can point at the agents your own team is building.

6 hoursIntermediateIncluded with the Gold Medal PassLimited seats

Overview

AI agents can now read your customers' records, change saved state, and call internal tools. Most of them are shipped with no more security testing than a few hand-typed jailbreak attempts. The hard part of testing them was never writing attack prompts. It is understanding what the agent is actually allowed to do, connecting a test harness to the real application, and proving whether an attack actually caused a forbidden effect or the model only said it did.

This hands-on course teaches you to do that work in a realistic lab of intentionally vulnerable agentic applications. You will start by attacking agents manually: direct and indirect prompt injection, cross-user data access, tool misuse, and system prompt leakage. You'll learn to separate instruction influence from authorization, and to verify findings against tool calls and saved state instead of the model's confident wording.

Then you'll put an AI testing agent to work. You'll direct it to recon a black-box application, build your own Promptfoo adapter, and run a live red team campaign. You'll repeat the process with application source, and judge every result as a confirmed violation, no observed violation, or inconclusive.

The afternoon turns the tables. You'll build an evaluation pipeline for agents you develop yourself: policy-based test cases, deterministic checks on tool calls and state, LLM judges that are calibrated before you trust them, and regression suites that show whether a fix closed the hole without breaking legitimate use.

One more thing… The course closes with a look at "abliterated" open-weight models (models with their refusal behavior removed) and how to use them as attack generators and pentest assistants against systems you are authorized to test.

When
Wednesday, February 24
Length
6+ hours w/ lunch being provided.
Format
Hands-on, lab-driven
Level
Intermediate

Gold price goes up after Nov 30

--days
:
--hrs
:
--min
:
--sec

$534.80 until November 30, plus Eventbrite fees. Then $668.50. All price steps

Course outline

  1. Module 1

    How AI agents fail

    • The agentic trust boundary: user → model → tool layer → data and saved effects
    • Prompt injection vs. authorization: why changing the tool choice should never grant permission
    • Excessive agency, sensitive data disclosure, and the OWASP LLM Top 10 (2025) as a coverage checklist
  2. Module 2

    Manual AI red teaming

    • Building a baseline from normal application behavior
    • Direct and indirect prompt injection through retrieved content
    • Paired identity testing: same object, same action, different caller
    • Verifying impact: tool events and saved state vs. what the model claims happened
  3. Module 3

    Agent-led red teaming, black box

    • Scoping, tooling, and budgeting an AI testing agent (and keeping target content from steering it)
    • Agent-driven recon: mapping sessions, identity, inputs, and errors from live traffic
    • Building and reviewing your own Promptfoo adapter
    • Plugins, strategies, and custom intents; running a live campaign
  4. Module 4

    Source-assisted testing

    • Tracing authorization checks and their call sites in application source
    • Deriving targeted test cases from code
    • Judging results: violation, no observed violation, or inconclusive
  5. Module 5

    Evaluation pipelines for agents you build

    • Turning a security policy into test cases and success criteria
    • Deterministic assertions on tool calls and state vs. model-graded rubrics
    • Calibrating LLM judges against labeled examples before trusting their verdicts
    • Separating attacker, target, and judge model roles
    • Measuring a fix: retained controls, regression runs, and honest reporting of rates
    • Running evals continuously in your development workflow
  6. Module 6 One more thing…

    Abliterated models for offensive testing

    • What abliteration is and how refusal removal works
    • Running uncensored open-weight models locally
    • Using them as attack generators and pentest assistants when commercial models refuse
    • Operational and ethical guardrails for authorized engagements

Key takeaways

  • Red team AI agents manually and prove impact with tool-call and state evidence, not model output
  • Direct an AI testing agent through recon, adapter building, and live Promptfoo campaigns while you keep the decisions
  • Test agentic applications in both black-box and source-available engagements
  • Build a repeatable evaluation pipeline for the agents your team ships, with calibrated judges and regression tests
  • One more thing… Use abliterated open-weight models as offensive tooling

Who should take this course

  • Penetration testers and red teamers adding AI applications to their engagements
  • Application security engineers responsible for LLM-powered products
  • Developers and ML engineers building agents who need security evaluations before (and after) shipping
  • Security leads who want a hands-on understanding of how AI agents are tested

Audience skill level

Intermediate. You should be comfortable on the command line, reading HTTP requests in browser developer tools, and reading basic JavaScript/TypeScript or Python. No prior AI or machine learning experience is required.

Student requirements

  • Basic web application testing concepts (requests, sessions, authentication vs. authorization)
  • Familiarity with JSON/YAML configuration files

What students should bring

  • A laptop with administrator rights and a modern browser
  • Specific hardware and software requirements will be announced before the training

What students will be provided with

  • Access to intentionally vulnerable agentic lab applications (black-box and source-available)
  • Student workbook, testing agent prompts, and adapter-building guides
  • Example Promptfoo configurations and evaluation pipeline templates to reuse at work
Bryson Loughmiller

Trainer

Bryson Loughmiller
Founder, Petracore

Bryson Loughmiller has spent more than 11 years in cybersecurity and tech across a range of security disciplines, at companies including Adobe, Podium, and Entrata. He founded Petracore to dive headlong into the AI whirlwind, building AI-powered software and running AI penetration testing engagements and trainings. He is also the co-founder of Ixway, an AI gateway that gives organizations transparency, observability, and control across workforce AI usage, model routing, and cost. Outside of work, Bryson enjoys photography, music, camping, building things, and sim racing.

Training day at past summits

Small rooms, laptops open, the instructor at the table with you.

Attendees on laptops during a hands-on training
A training room with mountain views through the windows
Training tables with laptops and a projected slide

Convince your boss

Get your training approved.

Our justification letter is ready to edit and send. It lays out the agenda, the value and a cost estimate you can adjust.

  • Sending a team? Email [email protected] about team pricing.
  • Paying by purchase order? We take purchase orders and other ways to pay, arranged by email.
  • Already have Silver or Bronze? Email [email protected] to upgrade to Gold.

Also on Wednesday

The trainings run side by side. Pick one for the day.

Bryce Kunz

Instructor

Bryce Kunz
Founder, Ripty

AI-gent Smith Training

A hands-on training exploring an emerging area of AI and cybersecurity. Full topic, prerequisites and learning objectives are coming soon.

6 hoursIncluded with the Gold Medal PassLimited seats