Resources, Security
AI agent security risks and how to reduce them
An AI agent reads untrusted content and then acts with real access. This guide sets out the main security risks, maps them to the OWASP lists and gives a mitigation for each.
Updated 20 November 2026. 9 minute read.
In short
- The OWASP Top 10 for LLM Applications 2025 ranks prompt injection first. It also lists sensitive information disclosure, supply chain, excessive agency, system prompt leakage and unbounded consumption.
- The OWASP Top 10 for Agentic Applications, published in December 2025 as the 2026 edition, adds agent-specific risks such as agent goal hijack, tool misuse, identity and privilege abuse, and memory poisoning.
- Prompt injection cannot be fully prevented today. The main defence is to limit what an agent can do: least privilege, separate untrusted content, and employee approval for high-impact actions.
- Give each agent its own scoped credentials, keep secrets out of prompts, log every action, and prepare an incident plan that includes pausing agents and revoking keys.
Why agents change the risk, and the OWASP lists to know
A chatbot produces text that an employee reads. An agent plans steps, calls tools and changes records. It also reads emails, web pages and documents that outsiders can write. That combination turns a wrong answer into a wrong action, which is why agent security is mostly about limiting access and impact.
OWASP Top 10 for LLM Applications 2025
The OWASP GenAI Security Project published the 2025 edition of its Top 10 for Large Language Model Applications in November 2024. It is the current edition. The ten risks are:
- LLM01 Prompt Injection
- LLM02 Sensitive Information Disclosure
- LLM03 Supply Chain
- LLM04 Data and Model Poisoning
- LLM05 Improper Output Handling
- LLM06 Excessive Agency
- LLM07 System Prompt Leakage
- LLM08 Vector and Embedding Weaknesses
- LLM09 Misinformation
- LLM10 Unbounded Consumption
OWASP work on agentic AI
The project's Agentic Security Initiative published Agentic AI: Threats and Mitigations in February 2025, a threat taxonomy for agents. Version 1.1 followed in December 2025. On 9 December 2025, OWASP also released the OWASP Top 10 for Agentic Applications, labelled the 2026 edition. Its risks are agent goal hijack (ASI01), tool misuse and exploitation (ASI02), identity and privilege abuse (ASI03), agentic supply chain vulnerabilities (ASI04), unexpected code execution (ASI05), memory and context poisoning (ASI06), insecure inter-agent communication (ASI07), cascading failures (ASI08), human-agent trust exploitation (ASI09) and rogue agents (ASI10).
Use the LLM Top 10 for the model layer and the Agentic Top 10 for what agents do with tools, memory and other agents. Most companies need both.
Direct and indirect prompt injection
Prompt injection happens when input changes the model's behaviour in ways its owner did not intend. In a direct injection, the attacker types the instruction into the agent. In an indirect injection, the instruction is hidden in content the agent processes, such as an email, a web page, a file or a tool result. NIST's taxonomy of adversarial machine learning, NIST AI 100-2 E2025, treats indirect prompt injection as its own attack class and notes that agents are exposed to both kinds.
Indirect injection is the bigger risk for agents, because an agent that reads the inbox reads whatever outsiders send. In the Agentic Top 10, this is the main route to agent goal hijack: content redirects the agent's objective, and the agent then uses its own tools against the company.
Mitigations
- Constrain the agent's role and tasks in its instructions, and define the output format you expect.
- Mark and separate untrusted content, such as emails and web pages, from instructions.
- Filter inputs and validate outputs, but do not rely on filters alone.
- Apply least privilege, so an injected instruction can only reach what the agent's job needs.
- Require an employee's approval before high-impact actions, such as payments, deletions or messages to outsiders.
- Test agents with adversarial prompts before launch and after changes.
OWASP notes that, given how models work, it is unclear whether prompt injection can be fully prevented. Plan as if some injections will succeed, and limit what they can do.
Excessive agency and tool misuse
Excessive agency (LLM06) is the risk that an agent takes damaging actions because it has too much power. OWASP names three root causes: excessive functionality (tools it does not need), excessive permissions (tools with more access than needed) and excessive autonomy (high-impact actions without a check). The Agentic Top 10 describes the agentic form as tool misuse (ASI02) and identity and privilege abuse (ASI03).
Mitigations
- Give each agent only the apps and tools its role needs, and remove the rest.
- Prefer narrow tools, such as send a reply to this ticket, over open-ended ones, such as run any shell command.
- Give tools read-only access unless the job requires writes.
- Act in the context of the right employee or a dedicated agent identity, never a shared admin account.
- Enforce approvals in the system that runs the tool, not in the prompt, so the model cannot talk its way past them.
- Rate-limit actions and set spending limits, which also addresses unbounded consumption (LLM10).
Where agents can write and run code, run that code in a sandbox with limited network access. This addresses unexpected code execution (ASI05).
Sensitive data disclosure, system prompts and memory
Sensitive information disclosure (LLM02) covers personal data, financial details, credentials and confidential business data leaking through model output. Agents widen the risk, because an injected instruction can ask the agent to send data out through any tool it holds, such as email or a web request.
System prompt leakage (LLM07) is related. OWASP's point is that system prompts are not secret and must not hold credentials or be the only security control. Memory and context poisoning (ASI06) is the reverse problem: false or malicious content saved in an agent's memory or retrieval store shapes later actions.
Mitigations
- Limit the data each agent can reach to what its role needs.
- Never put passwords, API keys or tokens in prompts or instructions.
- Control outbound channels: approve emails to new recipients and restrict web requests.
- Review what goes into shared memory, and record its source so poisoned entries can be traced and removed.
- Check model provider terms on data retention and training, and cover personal data in a data processing agreement.
Tools, supply chain and credentials
Supply chain risk (LLM03) covers third-party models, datasets, plugins and libraries. For agents, ASI04 extends it to tools, prompt templates, tool servers such as Model Context Protocol servers, and other agents loaded at run time. A compromised or impersonated tool can change agent behaviour or leak data.
Supply-chain mitigations
- Keep an inventory of the models, tools and connectors each agent uses, with their owners.
- Install tools only from sources you trust, pin versions and review changes before updating.
- Treat tool descriptions and tool output as untrusted input.
- Assess AI vendors as you would any supplier: security practices, data location and incident notice.
Credential handling
Agents need credentials to use apps. Identity and privilege abuse (ASI03) occurs when an agent inherits, reuses or escalates access it should not have. Good practice:
- Give each agent its own credential, never an employee's personal login.
- Scope each credential to the minimum permissions and rotate it on a schedule.
- Store secrets encrypted, outside prompts, and inject them only at the moment of the call.
- Protect the accounts of employees who manage agents with phishing-resistant sign-in, such as passkeys, or with two-step verification.
- Revoke credentials as soon as an agent is retired or its role changes.
Logging, monitoring and incident response
You cannot investigate what you did not record. OWASP lists logging and monitoring among the measures that limit the damage of excessive agency, and NIST's Generative AI Profile treats information security as one of its twelve risk areas. For agents, log each tool call, its inputs, its result, the approver and the cost, and keep the logs where the agent cannot alter them.
What to monitor
- Actions outside an agent's normal pattern, such as new recipients, bulk exports or unusual hours.
- Spending spikes and loops, which can signal a runaway or hijacked agent.
- Approval requests that are denied repeatedly.
- Changes to an agent's instructions, tools or credentials.
Incident response
Add agents to your existing incident plan rather than writing a separate one. Cascading failures (ASI08) and rogue agents (ASI10) show why speed matters: an error in one agent can spread to others.
- Contain: pause the affected agent and any agents that depend on it.
- Revoke and rotate the credentials it held.
- Preserve the activity log and export it for the investigation.
- Find the entry point, such as an injected email, a poisoned memory entry or a compromised tool.
- Assess notification duties, for example under the GDPR or NIS2.
- Fix the cause, tighten permissions or approvals, and test before resuming.
A practical checklist
- List every agent with its owner, purpose, data, apps, tools and credentials.
- Remove tools and permissions each agent does not need.
- Require employee approval for payments, deletions, external messages and other high-impact actions.
- Treat emails, web pages, files and tool output as untrusted, and keep them apart from instructions.
- Keep secrets out of prompts. Give each agent its own scoped, encrypted credential.
- Sandbox any code an agent writes or runs.
- Review what agents save to shared memory.
- Pin and review third-party tools and connectors, and assess AI vendors as suppliers.
- Set spending limits and rate limits per agent.
- Log every action and approval, keep logs tamper-resistant, and monitor for unusual behaviour.
- Add agents to the incident plan, with steps to pause agents and revoke credentials.
- Test with adversarial prompts before launch and after every significant change.
Least privilege and employee approval for high-impact actions do the most work. They limit the damage of risks you cannot fully prevent.
How Blentera helps
Blentera is a platform where companies run AI agents with employees in charge. Each agent has a role in the org chart and can use only the apps and tools it is given. Approval rules decide which actions need an employee's approval. Budgets warn at 80% and pause the agent at the limit. Credentials are encrypted with AES-256-GCM. Employees sign in with passkeys or two-step verification, with single sign-on on Business and user provisioning (SCIM) on Enterprise. Every action lands in an activity log that can be exported, and the platform is hosted in the EU. These controls support the steps above. Securing your agents remains your company's responsibility.
Questions
- Prompt injection, especially indirect injection through emails, web pages or files the agent reads. It ranks first in the OWASP Top 10 for LLM Applications 2025. The damage depends on what the agent can do, so least privilege and employee approval matter most.
- In a direct injection, the attacker types instructions into the agent. In an indirect injection, the instructions are hidden in content the agent processes, such as a document, a web page or a tool result.
- Excessive agency is LLM06 in the OWASP Top 10 for LLM Applications 2025. It is the risk that an AI system takes damaging actions because it has too many tools, too many permissions or too much autonomy.
- Yes. The OWASP GenAI Security Project released the OWASP Top 10 for Agentic Applications on 9 December 2025 as the 2026 edition. It sits alongside its Agentic AI: Threats and Mitigations taxonomy.
- Give each agent its own credential with the minimum permissions. Store it encrypted outside prompts and instructions, rotate it, and revoke it when the agent is retired or its role changes.
Sources
- OWASP Top 10 for LLM Applications 2025, OWASP GenAI Security Project
- LLM01:2025 Prompt Injection, OWASP GenAI Security Project
- LLM06:2025 Excessive Agency, OWASP GenAI Security Project
- OWASP Top 10 for Agentic Applications, OWASP GenAI Security Project
- OWASP GenAI Security Project releases Top 10 risks and mitigations for agentic AI security, 9 December 2025
- Adversarial Machine Learning: A Taxonomy and Terminology of Attacks and Mitigations (NIST AI 100-2 E2025)
- Generative Artificial Intelligence Profile (NIST AI 600-1), NIST
Related
This guide is general information, not legal advice. Check how the rules apply to your company with your own advisers.
Get started
Put your first agent to work in minutes.
Three steps, and your first agent is working for your company.
Step 1
Create an account
Sign up with your work email and invite your team. No sales call needed.
Step 2
Add your first agent
Create an agent, give it a role, and connect your own AI key and the apps it may use.
Step 3
Give it a task
Ask in Blentera or your chat app, and follow the work until it is done.