Comparisons of ways to test AI systems: AI red teaming, LLM application penetration testing, automated and human-led adversarial testing, agentic AI testing and AI risk assessment, against the OWASP GenAI Top 10 lists, MITRE ATLAS, NIST AI 600-1 and the EU AI Act.
Objective-driven adversarial testing of model behavior, misuse and harms against a security test of an LLM application and its surrounding web, API and data layers: scope, checklists versus scenarios, team skills, reporting, and which one the EU AI Act and customers ask for.
Red teaming aimed at an AI model's behavior, misuse and harms against red teaming aimed at an organization's networks, people and detection: the different technique catalogs (MITRE ATLAS versus ATT&CK), team make-up, success criteria, and when an AI-heavy organization needs both.
Scanners and model-driven attack generation that run thousands of adversarial prompts, against expert humans who craft context-specific attacks and judge harm: what NIST AI 600-1 says about each, coverage versus depth, cost, open-source tools such as garak and promptfoo, and how to combine them in release testing.
Testing agents that plan, call tools and act with delegated identity against testing chat or retrieval applications that mostly return text: the extra risks in the OWASP Agentic Top 10, MCP authorization issues, the test environment agents need, and how scope and duration change.
Governance-level identification and evaluation of AI risks, as ISO/IEC 42001, the NIST AI RMF and EU AI Act Article 9 require, against hands-on adversarial testing of a specific system: what each produces, which comes first, and how red team results feed the risk register.
An adversary technique knowledge base for AI systems against a ranked list of LLM application risks: structure, update cadence, how the 2026 LLM Top 10 maps to ATLAS, and which one to use for test planning versus reporting.
Risks of LLM applications against risks of autonomous agents that call tools, hold identities and talk to other agents: the two 2026 lists side by side, where they overlap, and which one should drive the scope of a test of your system.
Testing whether crafted or embedded inputs change an application's behavior against testing whether a model can be made to drop its safety rules: OWASP LLM01 definitions, direct and indirect injection, ATLAS technique IDs, and which matters more for a business app.
Reviewing a Model Context Protocol server against the MCP security best practices against a conventional API penetration test: token passthrough, confused deputy and local server risks, overlap with the OWASP API Top 10, and how to scope one engagement for both.
Tell us what you're working on, whether it's a deployment, an audit, a security test or a cyber range. You'll speak with an engineer who can help you scope it.