Agentic AI Security
Agentic AI security is protecting AI agents, systems that plan and take actions on their own using tools, data and other software, so they can't be hijacked, misuse their access or cause harm through mistakes.
Also called: AI agent security, autonomous agent security, OWASP Top 10 for Agentic Applications, agent goal hijack
What is agentic AI?
A chatbot answers questions. An AI agent gets a goal, works out the steps and carries them out: searching the web, reading files, using business applications, writing and running code, sending messages and sometimes handing work to other agents. Examples include coding agents, AI that handles customer service tickets end to end, and assistants that book travel or process invoices.
That autonomy is the point, and also the risk. Agentic AI security is about making sure agents do what their owners intend, and nothing else.
Why agents are different
- They act, not just answer. Mistakes or manipulation turn directly into real-world actions.
- They read untrusted content as part of their job, which exposes them to indirect prompt injection.
- They hold credentials and permissions, effectively becoming a new kind of user that needs identity and access management.
- They chain steps together, so one bad decision can cascade, and humans may not see each step.
- They may remember information across sessions, which attackers can try to corrupt.
Key risks
In December 2025, OWASP published its first Top 10 for Agentic Applications, built with more than 100 experts. Its risks include:
- Agent goal hijack: an attacker redirects the agent's objective, often through content it reads, rather than by breaking any code.
- Tool misuse: the agent uses a tool it's allowed to use, but in a harmful or unintended way.
- Identity and privilege abuse: agents with broad or borrowed credentials.
- Memory and context poisoning: planting false information the agent relies on later.
- Insecure communication between agents, and cascading failures across multi-agent systems.
- Human-agent trust exploitation: people approving whatever an agent proposes.
- Rogue agents acting outside their intended scope.
How to deploy agents safely
- Apply "least agency": give each agent only the tools, data and permissions its task needs. See excessive agency.
- Give agents their own identities with scoped, short-lived credentials, never a shared admin account.
- Keep a human in the loop for high-impact actions, and make approval prompts clear enough to review properly.
- Treat everything an agent reads as untrusted.
- Log every action and monitor for unusual behavior, with a way to stop an agent immediately.
- Sandbox code execution and limit network access.
- Test agents adversarially before launch with AI red teaming.
Related terms
Sources
- OWASP Top 10 for Agentic Applications for 2026 (OWASP Gen AI Security Project)
- LLM06:2025 Excessive Agency (OWASP Gen AI Security Project)
Drafted with AI assistance, then edited and fact-checked by a person. How we use AI.
