AI agents are shipping fast. QA is catching up slowly. Most teams are still running the same LLM evaluation suites they built for chatbots, then attaching them to systems that can write to databases, trigger deployments, and call external APIs. The gap between what those evals catch and what agents can actually do in production is where the real risk lives.
OWASP has named that gap. In 2026, the Open Worldwide Application Security Project published its first Top 10 for Agentic Applications, a peer-reviewed framework identifying the most critical security risks facing autonomous AI systems. It covers territory that no standard LLM benchmark touches.
Ten risks your current eval suite probably misses
The OWASP Top 10 for Agentic Applications introduces a new vocabulary for agentic failure modes. A few stand out as immediately relevant for QA and product teams.
Agent Goal Hijack (ASI01) is the agentic version of prompt injection. An attacker embeds malicious instructions in a ticket, a wiki page, or a retrieved document. The agent reads it, treats it as instruction, and acts on it. Your eval probably tests adversarial prompts in isolation. It almost certainly does not test what happens when the adversarial content arrives through a tool call.
Memory and Context Poisoning (ASI06) targets the persistent memory stores that give agents continuity across sessions. Corrupted context shapes future behaviour in ways that are hard to detect and harder to trace. Static benchmarks cannot catch this because they reset state between runs.
Cascading Agent Failures (ASI08) describes how a small error in one agent propagates through connected systems. In a multi-agent architecture, a single compromised component can affect every downstream decision. Evals that test agents in isolation will never surface this class of failure.
Research published in May by Help Net Security adds another dimension: "confused-deputy" attacks, where an authorised system is manipulated into misusing its own privileges. The research found that prompt-only defences are brittle under adversarial conditions and advocates a strict "propose-commit split" architecture, where the model can reason and draft but cannot execute writes without a separate policy gate and human approval step.
What to test first
If your team is shipping agentic features, three test areas should move to the top of your backlog.
First, test goal persistence under indirect injection. Feed the agent adversarial instructions embedded in retrieved content, not in the direct prompt. Observe whether its objective shifts.
Second, stress-test memory across sessions. Introduce corrupted or contradictory entries into the agent's memory store and evaluate whether downstream reasoning degrades.
Third, map your tool permissions. Most agents have more access than they need. Applying least-privilege to tool grants is one of the highest-value controls available before any eval is written.
OWASP's framework is a starting point, not a complete test plan. But it is the most authoritative signal yet that agentic evaluation requires its own discipline, separate from LLM testing and separate from traditional software QA.
If you are building that test plan, EvalSpec is designed to help.
Try EvalSpec free - evaluate your AI system across all 5 dimensions. No code required.
Start your free evaluation