OWASP published three separate AI security landscapes in Q2 2026. That is not a coincidence. It is the security community formally acknowledging that agentic AI systems require a different testing model to traditional LLM applications. If your AI testing strategy was designed for prompts and responses, it may not be ready for what agents can do.
Why agentic risk is different
Standard LLM security focuses on what goes in and what comes out. Prompt injection filters, output sanitisation, access controls. These are well understood and widely implemented.
Agentic systems operate differently. An AI agent does not respond to a single prompt and stop. It plans across multiple steps, calls tools, coordinates with other agents, and makes decisions without human review at each point. The risks emerge during execution, not at the input or output boundary.
Multi-turn reasoning chains can drift. A tool called mid-task can behave unexpectedly. A sub-agent in a multi-agent pipeline can be compromised. None of these are detectable through static analysis or simple guardrail checks.
The OWASP Top 10 for Agentic Applications 2026, developed with over 100 security researchers and practitioners, names Agent Goal Hijacking as its top risk. Categorised as ASI01, this is when an agent's objective is redirected mid-task, by a malicious input, a compromised tool, or an adversarial instruction passed through the pipeline. As agents are connected to real systems with real permissions, the consequences of a hijacked objective grow significantly.
Red teaming is now a discipline, not a milestone
On 10 April 2026, OWASP published its first dedicated Red Teaming Landscape for AI and agentic systems. Its core message is direct: the thing you are testing keeps changing.
An agent's behaviour shifts as its underlying model is updated, its tool set expands, or its operational context changes. A red team exercise conducted before release does not account for any of this. The landscape calls for continuous adversarial testing throughout the development lifecycle, paired with runtime enforcement at the agent layer rather than the API level.
This shifts red teaming from a pre-deployment checkbox to an ongoing practice. It belongs in your testing cadence alongside automated evaluation runs.
Three things to put in place now
If your team builds or deploys agentic AI, three capabilities matter.
- Inventory. Know which agents exist, what tools they can access, and what permissions they hold. You cannot test what you cannot see.
- Runtime controls. API-level guardrails catch a narrow class of problems. Enforcement needs to sit at the agent layer, where decisions are actually made.
- Continuous testing. Schedule adversarial testing as a regular activity, not an exception. Update your threat models when your agents change.
OWASP's Q2 2026 publications are a practical starting point. The Top 10 for Agentic Applications and the Red Teaming Landscape both offer actionable frameworks, not just risk taxonomy.
EvalSpec does not evaluate agent behaviour, tool calls or runtime controls today. It gives you structured, repeatable evaluation of the responses your AI system returns, including prompt injection and other safety cases mapped to the OWASP LLM Top 10.
Try EvalSpec free - evaluate your AI system across all 5 dimensions. No code required.
Start your free evaluation