
The OWASP GenAI Security Project has published the Top 10 for Agentic Applications — a list of the ten critical risks for agent-based applications. The document was developed by more than a hundred experts and underwent peer review.
Why a Separate List Is Needed
A standard language model receives a question and returns text. An agent does something entirely different: it plans, decides and acts. An agent can read emails, query a database, draft a contract, call an external API and send the result to a colleague — without a person approving each step.
That difference is what justifies a separate risk list. When a system only answers, the maximum damage is a wrong answer. When it acts, the damage is a wrong action — and sometimes an irreversible one.
The Ten Areas
- Planning — hijacking the agent's goal
- Tool use — unintended or malicious tool invocation
- Identity — an agent acting with someone else's permissions
- Supply chain — third-party components and libraries
- Code execution
- Memory — poisoning the context the agent remembers
- Inter-agent communication
- Cascading failure — one agent's error that spreads
- Human–agent trust
- Rogue agents
The Practical Recommendations
Three recommendations recur: behavioral monitoring that understands what the agent does and not just what it says; a human in the loop for critical actions — payments, data deletion, permission changes; and red team exercises explicitly targeting prompt injection and tool misuse.
It is worth cross-referencing this with a finding from the IBM report from the same month: more than 50% of organizations already run agents for threat detection, but only 18% for vulnerability management. The agents are already inside — the question is whether anyone is mapping their risk.
What This Means for Your Organization
- Make a list of every AI agent running in your organization, including ones embedded in existing tools without an explicit decision.
- For each agent — check which permissions it operates with. The simple rule: an agent should not hold broader permissions than the employee it replaces.
- Define in advance which actions require human approval. Payments, deletion and permission changes are a reasonable starting point.
- Remember the risk of cascading failure: two agents talking to each other can amplify an error instead of correcting it.
