In short
- Holistic AI already red teams AI agents, not just models. This release makes it much smarter.
- It now learns what each agent can do and tailors the attacks to that agent.
- Attacks adapt as the conversation goes, the way a real attacker would.
- Every finding comes with clear evidence and a recommended fix.
- It is built on research that placed Holistic AI in the Top 10 of OpenAI's global red teaming hackathon.
Built on research that proved the gap
In 2025, OpenAI ran a global red teaming hackathon for its open weight model, GPT-OSS-20B. More than 600 teams took part. Holistic AI, working with University College London, finished in the Top 10.
Our research showed something many teams miss. A model that looks safe on its own can fail once it runs inside an agent.
In one test, an attack failed every time against the model alone. Inside an agent, the same attack worked 67% of the time.
That is why the Holistic AI Governance Platform has tested agents, not just models, as part of agentic red teaming. You can read the full research in Mind the Gap. This release takes it further.
Why agentic AI governance needs agentic red teaming
AI agents do more than answer questions. They look up records, use tools and take actions.
That changes what can go wrong. An agent might share someone else's data, approve something it should not, or give in when a user pushes hard.
Governance teams need to know one thing: does the agent stay inside its boundaries? Rules like the EU AI Act and frameworks like NIST AI RMF expect testing evidence, not assumptions.
Agentic red teaming gives you that evidence for the systems that now matter most. It is part of how an AI governance platform turns AI risk into something you can measure and manage.
What's new in Agentic Red Teaming 2.0
Agentic red teaming is already part of the Holistic AI Governance Platform. Version 2.0 upgrades every step of it, from understanding the agent to reporting what to fix. The result is testing that behaves more like a real attacker and gives your team clearer answers.
Here is what has been upgraded.
Upgraded: It learns the agent first
Before any attack, the platform has a normal conversation with the agent. It learns what the agent is for, what data it uses and what it can do.
That builds a profile of the agent and its likely weak spots. Tests that do not apply are skipped, so time goes where the risk is.

Upgraded: Attacks tailored to your agent
Instead of the same tests for every system, the platform suggests attacks that fit this agent. Each one comes with a plain reason why it applies.
You stay in control. A person reviews the plan and chooses what to run.

Upgraded: Attacks that adapt like a real attacker
Real attackers do not give up after one refusal. They change their approach.
Our attacks do the same. They respond to what the agent says and keep testing the boundary, so you see how the agent holds up under real pressure, not just a single prompt.

Upgraded: Findings you can trust
Every result is reviewed before it becomes a finding. A finding always comes with the evidence from the conversation, so nothing rests on a guess.
The report shows what happened, how often, how serious it is and what to fix.

Upgraded: Coverage for the risks agents create
- Data extraction: can the agent reveal information the user should not see?
- Tool abuse: can someone push the agent into an action outside its authority?
- Bias: do similar requests get different outcomes because of who is asking? See AI bias testing.
- Sycophancy: does the agent drop a correct answer or a safe decision when a user pushes back?
These sit alongside our existing tests for jailbreaking, prompt injection and toxicity.
How it fits into AI governance
Red teaming is one step in a connected governance process, not a one off test.
- Know what to test. AI discovery finds the agents in your organisation, including ones nobody registered. Each one lives in the AI inventory.
- Test before release. AI red teaming shows where an agent's boundaries hold and where they break, as part of the Protect stage.
- Fix and retest. Each finding points to what needs fixing. Teams fix it and run the same test again.
- Protect in production. Guardian Agents watch agents at runtime and can block risky actions.
- Prove it. Test results and fixes become evidence for audits and regulators.
Why Holistic AI
Proven on the world stage. Our agentic red teaming research placed in the Top 10 of more than 600 teams in OpenAI's global hackathon.
Testing is in our roots. Holistic AI started in AI auditing. Bias, robustness and adversarial testing have been part of the platform from the start. Read more in Red Teaming at Holistic AI.
One platform from discovery to runtime. Many tools focus on one part of the job: testing, monitoring or policy. The Holistic AI Governance Platform brings all three together, so a finding in testing connects to the same agent in your inventory and the same controls in production.
Recognised by analysts. In Gartner's 2026 Critical Capabilities for AI Governance Platforms, Holistic AI had the highest score of 13 vendors for AI Risk and Compliance.
At a glance
Part of the Holistic AI Governance Platform
Agentic red teaming is one part of the Holistic AI Governance Platform. It works with everything else in it: AI discovery and inventory across Identify, testing across Protect, and runtime enforcement across Enforce.
Get a demo
See how agentic red teaming and the full AI governance platform work on your own AI agents. Request a demo.


