AgentMill

Securing AI agents: 10 risks

As of 15 November 2026 · Author: Tarkan Yavas, Cybermind Academy

AI agents read content, make decisions and act with real permissions. That shifts the risk: a chatbot giving a wrong answer costs time. An agent acting wrongly can leak data or trigger processes.

This guide follows the OWASP Top 10 for Agentic Applications 2026 (published 9 December 2025). The identifiers ASI01 to ASI10 come from OWASP, the explanations and countermeasures are our own summary. The OWASP list for LLM applications 2025 is a useful complement, for example on prompt injection, sensitive information disclosure and excessive agency.

The ten risks

ASI01 1. Agent goal hijack

What it is: An attacker redirects the agent’s goal or plan through content it reads, such as an email, a document or a web page. The agent cannot reliably tell text from instructions.

What helps: Treat external content as untrusted, separate instructions from data, require approval for actions with external effect, monitor deviations from the task.

ASI02 2. Tool misuse

What it is: The agent uses legitimate tools in ways that were not intended: it deletes, sends or exports more than planned.

What helps: Define allowed tools via allowlist, restrict parameters, set limits, require confirmation for destructive actions.

ASI03 3. Identity and privilege abuse

What it is: Agents inherit excessive permissions, share accounts or tokens, or act under a human’s identity.

What helps: Dedicated identity per agent, least privilege, short-lived credentials, no use of private or shared accounts.

ASI04 4. Agentic supply chain

What it is: Third-party tools, plugins, interface servers, prompts or models are loaded at runtime and may be tampered with.

What helps: Keep an inventory of all components, allow only vetted sources, pin versions, review changes.

ASI05 5. Unexpected code execution

What it is: Agents generate or run code, such as scripts or shell commands. An attack can lead to remote code execution.

What helps: Run code in a sandbox without access to production systems, review code, require approval to execute, restrict network access.

ASI06 6. Memory and context poisoning

What it is: Poisoned memories, knowledge bases or retrieved documents influence decisions persistently.

What helps: Validate sources, separate memory per user and tenant, set expiry, be able to inspect and reset content.

ASI07 7. Insecure inter-agent communication

What it is: Messages between agents are not protected and can be forged or altered.

What helps: Authenticate agents mutually, sign messages and validate against a schema, no implicit trust relationships.

ASI08 8. Cascading failures

What it is: A fault or attack spreads across several agents and systems and amplifies.

What helps: Isolate components, add limits and circuit breakers, provide a kill switch, test failure scenarios.

ASI09 9. Human-agent trust exploitation

What it is: People trust convincingly worded agent statements and approve harmful actions (automation bias).

What helps: Design approvals with context and sources, show uncertainty, train staff, apply four-eyes for critical actions.

ASI10 10. Rogue agents

What it is: Agents deviate from their task, for example through misalignment or compromise, and act on their own.

What helps: Monitor behaviour continuously, log actions, keep a kill switch ready, review agents regularly.

First steps

  1. Inventory all agents and assistants, with purpose, data and owner.
  2. Limit permissions per agent to what is needed and use dedicated accounts.
  3. Allow actions with external effect only after human approval.
  4. Set up logging and a kill switch.
  5. Run a risk review before the agent works with sensitive data.

To help you, there is the checklist with 18 checks. For an in-depth assessment of your agents we offer the AI agent risk analysis.

Sources

This article is a professional assessment and not legal advice. As of 15 November 2026.

Let’s talk about your project.

Request a conversation