Red-team and harden your AI agent team
You've built a single agent. You've built a team. Now put that team through what a real security review looks like: turn on production guardrails, then try to break it with five real attack techniques, and see whether your defences hold. Still no code. Your work stays in this browser and is never sent anywhere.
Listen to this briefing
Hardening and Red-Teaming an AI Agent Team
New here? Start with the first two tools
This tool builds on Advanced AI Agent Builder, which itself builds on Build Your Own AI Agent. If you haven't built a Supervisor-and-Specialist team before, start there first: this tool assumes you already know what a Supervisor, a Specialist, and a checkpoint are.
What's different here
A team that plans, runs, pauses for approval, and recovers from an ordinary tool failure.
The same kind of team, hardened with production guardrails, then deliberately attacked with 5 real techniques to see which guardrails hold.
What you will do: 5 steps
- 1Choose a team. Load a production-style team, the same shape you'd build in the Advanced tool.
- 2Harden it. Turn on the guardrails a real deployment would need, before anything goes live.
- 3Red-team it. Pick a real attack technique and watch it play out against your team, guardrails and all.
- 4Respond to the incident. Contain it, decide who needs to know, and log the root cause.
- 5Get your governance record. A structured record of every guardrail, every attack and every outcome, ready as the starting point for an auditor.
Your guardrails and attack results are kept in this browser as you go, so you can leave and come back. Press Start over to clear them.
Step 1 of 5: Choose a team to harden
Choose a team to harden
Pick a production-style team, the same shape as the Advanced tool's templates (a Supervisor plus two Specialists), so you can jump straight to hardening it.
Real hardening reviews start with a team that's already been designed and approved, not a blank page. These three are the same shape as the production deployments most companies stand up first.
Harden your team
Before anything goes live, real teams turn on a set of production guardrails. Toggle the ones you'd want running for this team. You'll see exactly what happens without them in the next step.
Every guardrail above is a real, published defence pattern, not a made-up rule. Leave one off, and the matching attack in the next step will get through.
Red-team it
Pick an attack. Each one targets exactly one of the guardrails from Step 2. Watch what your team does, and whether the guardrail you set stops it.
Attack simulation (not a real AI)
This is a practice tool that teaches the ideas behind agent hardening using simple, built-in rules. It does not call a real AI model, and no real attack is performed. Everything you see stays in your browser.
Every attack here is a simplified version of a real technique from the guardrail patterns in the Agentic AI Pattern Catalogue and the OWASP Top 10 for Agentic Applications: the same categories real red teams test for before launch. Each one runs on its own against a single guardrail; a real red team also chains multiple techniques together in one pass, which this tool does not simulate.
Respond to the incident
Test at least one attack in Step 3 to see a response walkthrough here.
Who should know about this?
How a team responds in the first few minutes after an attack, caught or not, is what regulators and auditors check. A good response turns a bad moment into a logged, fixable gap.For grown-ups: this maps to Pattern HIL-END-01 · Escalation & Handoff to Human in the Agentic AI Pattern Catalogue.
Your expert governance record
The same idea as the last two tools' records, scaled up again: this time with a residual risk rating and a full guardrail-and-attack log.
This record is simplified for learning. It cites real, published guardrail patterns from the Agentic AI Pattern Catalogue: 61 patterns for autonomous agents, mapped to real regulatory standards including Singapore's IMDA Model Governance Framework, ISO/IEC 42001, the OWASP Top 10 for Agentic Applications, and the US NIST AI Risk Management Framework. Finishing this tool does not mean a real agent team is certified, secure, or compliant. For a real organisation, use the full ISO 42001 AIMS Readiness Checklist and TRACE Agent Evaluation tools, or read the standards reference in full.
Four ideas this tool adds
You already know Supervisor, Plan, Checkpoint, and Recovery from the Advanced tool. Hardening a team for production adds four more ideas.
Guardrail
A specific, always-on control that blocks one category of attack, like screening tool output, or redacting secrets before they leave the team.
Attack surface
Everywhere your team could be tricked, fooled, or exploited: a poisoned webpage, an over-eager task, or another agent's bad guess.
Red-team
Deliberately attacking your own system before someone else does, so you find the gaps while they're still cheap to fix.
Residual risk
What's left over after your guardrails are in place: never zero, but a number you can see and manage.
Five attacks, each aimed at one guardrail
The pairings and rating rule the tool uses. It is a practice simulation with built-in rules: no real AI model is called and no real attack is performed. Leave a guardrail off and its matching attack gets through.
This is the tool I wish more people built before, not after, something goes wrong. Every attack in here, the poisoned search result, the scope creep, the memory that should have been wiped, is a simplified version of something that has happened to a real agent somewhere. Turning on a guardrail feels abstract until you watch the matching attack walk straight through the gap you left open. That's the whole point: find the gap here, where the cost of being wrong is nothing.
Last updated 2 Oct 2026
More in Agent labs
- Build Your Own AI AgentHow does an AI agent actually work, and what makes one safe to launch?
- Advanced AI Agent BuilderWhat changes when agents work as a team?
