Agentic AI and security: seven questions to ask before you deploy
What changes when an AI system stops answering and starts acting? The short answer: everything about how you need to think about risk. As agentic AI moves from pilots to production, the OWASP Top 10 for Agentic Applications gives developers and product owners a structured way to ask the right questions before deployment.
A generative AI application produces an answer or a recommendation. An AI agent goes further – it can receive a goal, decide which steps are needed, retrieve information, choose and use tools, interact with other systems, and carry out actions on its own. That shift from advising to acting is why agentic AI needs a different security conversation, than many teams are used to from AI assistants like ChatGPT andCopilot.
In December 2025, the OWASPGenAI Security Project published the OWASP Top 10 for Agentic Applications2026, a peer-reviewed framework built with more than 100 security experts to map the risks unique to autonomous, tool-using AI systems. It covers ten risk categories, from prompt injection and tool misuse to identity abuse, memory poisoning and cascading failures across connected agents.
For teams building or evaluating agentic AI, that framework boils down to seven practical questions worth asking before an agent is given the ability to act.
Seven questions to ask before you give an agent the ability to act
- What are we building? A shared understanding of what makes a system "agentic" in the first place.
- How autonomous should it be? Autonomy should be earned deliberately, not granted by default.
- What can influence it? Prompts, data sources and connected tools can all be manipulated.
- What can it access and do? Permissions and identity controls define the blast radius.
- What data and context can it retain? Memory and context can be poisoned or corrupted.
- When must a human take over? Some decisions still need a human checkpoint.
- Is it safe enough for production? A practical readiness check before go-live.
Over the next two posts, we'll go deeper into each of these questions – starting with autonomy, influence and access, then moving on to memory, human oversight and production readiness.
If your team is building or shipping agentic systems, our new course, Building safe AI agents, walks through all seven in practical detail, mapped directly to the OWASP framework.
