Introducing Giskard's Autonomous Agent Audit (AAA)

Hello there,

Now that the "Black October" horror show is over for software security, let's think about the future: securing AI itself.

In July 2026, AI agents under test escaped their sandbox and broke into Hugging Face. In September 2026, an AI hallucination nearly triggered a US military operation.

For security professionals, there are two lessons to derive from these 2 incidents:

1. Agents are now very capable autonomous attackers. Lesson: human red-teaming is not enough; defenders should put agents to work to secure their systems.

2. Agents can cause harm with no attacker (human or AI) behind them. Lesson: your tests must look for more than attacks, but also for safety and hallucination risks.

For developers building AI agents, there is a third lesson: your customers, your CISO and your investors will soon ask you to prove your agent is safe.

To solve that problem, we are excited to announce the next generation of red-teaming for AI systems: Giskard's Autonomous Agent Audit (AAA). Yes, we know: AAA is also a credit rating, a battery size and a kind of video game. We decided this was a good sign.

What's new about Giskard's Autonomous Agent Audit?

AAA is a harness of AI agents that audits your AI agent end-to-end.

This controlled swarm of agents does automated recon to understand how your agent works and what tools & accesses it has, models the risks of what could hurt your business, collaborate in adaptive waves to test the target AI agent (with the ability to chain testing techniques), ranks every finding by agent-specific severity, and last but not least, it generates reproducible tests for each finding and recommendations of remediations.

Test coverage includes the OWASP Top 10 for LLM and for Agentic applications, and goes beyond them. And to reassure you: every agent in the swarm is sandboxed, supervised and scoped to test only your AI system, and every exchange is logged.

What that gives your team:

• Business risks, not only security exploits. Wrong advice, false commitments and out-of-scope answers are tested alongside injection and data leaks.

• A detailed audit report with evidence and recommended fixes, readable for security & compliance stakeholders and actionable for the AI engineering team

• An external attestation, not just an internal report. Once your team has fixed the findings and the re-test passes, Giskard issues an attestation. You can share it with your customers, procurement and investors.

• Reproducible tests. The tests generated by the audit become a test suite you can replay at every model, prompt or tool change, in your CI/CD pipelines.

Human red-teaming depth, in a fraction of the time. Embedded in your team's workflow.

How does it compare to other AI red-teaming tools? We benchmarked Giskard's AAA against NVIDIA Garak, OpenAI Promptfoo, and Claude. Here are the results below:

Full benchmark details: https://www.giskard.ai/knowledge/automated-ai-red-teaming-benchmark

Want a demo? You can watch this short video, or watch the replay of our live session with a full demo on AAA and a technical presentation by our co-founder Jean-Marie John-Mathews.

📚 Want to learn more? Read our latest technical articles

AI red teaming vs. penetration testing: why pentest methodology breaks on LLM agents

A pentest works because a machine can verify whether an exploit succeeded. An LLM agent removes that signal: instructions and data share one context window, so prompt injection is a property of the architecture, not a bug you can patch like SQL injection. The article explains why verdicts become statistical and business-specific, what stays in pentest scope, and the seven components an AI red teaming capability needs.

→ Read the article

Harm-first security: why we should start from the harm, not from the techniques

In the OWASP GenAI working group, practitioners ranked prompt injection first and misinformation last, while incident records show the opposite. At Giskard, we argue for reversing the order: name the harm, then ask who or what can cause it, and let techniques follow. Drawing on NotPetya, NIS2 and DORA, we shows why AI scanners and prompt batteries are "advanced spell checkers", useful but never a risk model.

→ Read the article

What's next: Autonomous Agent Audits in self-service

We are building a fully self-service version of AAA, so that teams at startups or small-to-medium businesses can get their AI agents audited quickly, and test Giskard for themselves.

No need for a dedicated security team or human pentesters. Direct the Giskard AAA at your AI agent, get the report and the replayable test suite, correct your agent, and get your attestation.

We're opening a closed beta to a small number of teams. Who’s eligible:

• Companies founded in the last 5 years

• Less than 500 employees

If you need to get you AI agent audited and want early access, ask us to join the beta:

Best,

Alex Combessie & Jean-Marie John-Mathews, co-founders of Giskard AI

Built in Europe 🇫🇷 with Trust, Privacy & Safety