The OWASP Top 10 for LLM Applications, mapped to email

The OWASP Top 10 for LLM Applications is a community list of the most critical security risks in software built on large language models. The current edition is 2026, and prompt injection is LLM01, the top risk. For an agent with an inbox every item takes a concrete form, because email is where untrusted text arrives.

Updated

What is the OWASP Top 10 for LLM Applications?

The OWASP Top 10 for LLM Applications is a ranked list of security risks for applications that use large language models. It is written by the OWASP GenAI Security Project, a volunteer community of security practitioners.

The current edition is 2026, published in August 2026. Prompt injection is LLM01, the top risk, as it was in 2025.

The 2026 edition keeps the ten risk areas of 2025, renames one and changes the order. Older reports use the 2025 numbers, so this guide gives both.

What are the ten items in the 2026 list?

The ten items of the 2026 list, in order, with the number and name each had in 2025.

2026NameIn the 2025 list
LLM01Prompt InjectionLLM01 Prompt Injection
LLM02Sensitive Information DisclosureLLM02 Sensitive Information Disclosure
LLM03Excessive AgencyLLM06 Excessive Agency
LLM04Supply ChainLLM03 Supply Chain
LLM05Data and Model PoisoningLLM04 Data and Model Poisoning
LLM06Unbounded ConsumptionLLM10 Unbounded Consumption
LLM07MisinformationLLM09 Misinformation
LLM08Hidden Context ExposureLLM07 System Prompt Leakage
LLM09Vector and Embedding WeaknessesLLM08 Vector and Embedding Weaknesses
LLM10Improper Output HandlingLLM05 Improper Output Handling

A full identifier carries the year, as in LLM03:2026. When a report cites a number above LLM02, check which edition it means.

What does each item mean when an agent has an inbox?

Each item has a form that arrives or leaves by email. The sections follow the 2026 order.

LLM01:2026 Prompt Injection

Prompt injection reaches an email agent as a message it did not choose to read. The message carries an instruction, in plain sight or hidden from the human, and the model may follow it. This is indirect prompt injection.

Control: screen mail before the agent reads it, and give the model what a human could see.

LLM02:2026 Sensitive Information Disclosure

Sensitive information disclosure is the risk that an agent repeats what its mailbox holds: contracts, invoices, personal data, sign-in codes. It can be steered to put them in a reply, a summary or a link. See LLM data exfiltration.

Control: limit which mail the agent can read, keep sign-in codes away from it and screen what it sends.

LLM03:2026 Excessive Agency

Excessive agency is the gap between what a mail agent needs to do and what it is able to do. OWASP's own example is a mail assistant that needs to read, has a tool that can also send, and is told by an injected email to forward what it finds. OWASP names three causes: excessive functionality, excessive permissions and excessive autonomy.

Control: read or draft access where that is enough, and human approval for sends that matter.

LLM04:2026 Supply Chain

Supply chain risk for an email agent sits in what it depends on: a model, a framework, connectors and tool servers. A compromised or careless component stands between the mailbox and the model.

Control: use few components, pin their versions and give each one a key limited to the inbox it serves.

LLM05:2026 Data and Model Poisoning

Data poisoning reaches an email agent through mail that is stored and reused. A planted message can sit in a thread history, a memory store or a knowledge base and shape answers long after it arrived.

Control: screen mail before it is stored, and keep held mail out of anything the agent indexes.

LLM06:2026 Unbounded Consumption

Unbounded consumption follows from one fact: anyone can send mail, so anyone can make an agent work. A flood of messages, very long threads or a reply loop between two agents can run up model costs or stall real work.

Control: rate limits, size limits, loop detection and a spending cap on what one sender can trigger.

LLM07:2026 Misinformation

Misinformation from an email agent is a false statement sent with confidence: a wrong payment date, an invented policy, a summary that reverses a message's meaning. An injected email can cause it on purpose.

Control: keep a human on replies that commit the company, and keep the original message one step from any summary.

LLM08:2026 Hidden Context Exposure

Hidden context exposure happens by email when a sender gets an agent to reveal what it was given out of sight: its system prompt, its rules, the tools it can call. In 2025 this was System Prompt Leakage, and the 2026 edition widened it.

Control: keep secrets out of the prompt, assume it can be discovered, and enforce the rules that matter outside the model.

LLM09:2026 Vector and Embedding Weaknesses

Vector and embedding weaknesses appear when an agent indexes mail so it can search by meaning. If the index mixes mailboxes, or takes in unscreened mail, one sender's text can surface in another reader's answers.

Control: keep each inbox's index separate, and index delivered, screened mail and nothing else.

LLM10:2026 Improper Output Handling

Improper output handling is the risk in what happens to the model's writing next: a client renders it, a tool acts on it, a recipient reads it. A link, an image address or a command in that output can do harm if nothing checks it.

Control: treat model output as untrusted, and screen outgoing mail before it leaves.

Which items matter most for an email agent?

The top three of the 2026 list do most of the damage together: Prompt Injection, Sensitive Information Disclosure and Excessive Agency. Injection is how the attacker gets in. Agency is what the agent can be made to do, and disclosure is what leaves.

OWASP moved Excessive Agency from sixth in 2025 to third in 2026. The 2026 edition also says no reliable way to prevent prompt injection exists today, so the system around the model must limit what a fooled model can do.

OWASP also publishes a Top 10 for Agentic Applications, for agents that plan and act on their own. See AI agent security.

Which items does Trotta Mail address?

Trotta Mail addresses the items that pass through the inbox. It does not replace controls inside your agent or your model.

Item (2026)What Trotta Mail does
LLM01:2026 Prompt InjectionScreens each message before delivery, attached documents included, and holds dangerous mail for a human. An agent's key reads the safe view, with hidden text removed and instructions aimed at a model replaced by a note.
LLM02:2026 Sensitive Information DisclosureSeals sign-in codes and reset links in a human's inbox from every API key. Screens outgoing mail for secrets and sensitive data.
LLM03:2026 Excessive AgencyAn agent's key cannot release held mail or loosen its own protection. In a human's inbox an agent starts with drafts, not sends. Approval can be required for a first contact or for every send.
LLM10:2026 Improper Output HandlingScreens what an agent sends before it leaves. A reply steered to an address the inbox has never written to waits for a human.

The rules are in the security model. Supply chain, stored data, misinformation and consumption limits inside your agent remain yours to control.

No filter catches every injection. A reworded request with no technical tell can still be delivered, so OWASP's other mitigations still apply: least privilege, human approval and regular testing.

Questions

What is LLM01 in the OWASP Top 10 for LLM Applications?

LLM01 is Prompt Injection, the top risk in the OWASP Top 10 for LLM Applications, in both the 2026 and 2025 editions. It covers input that changes a model's behavior in ways its developer did not intend, whether typed directly or hidden in content the model reads.

What is the current edition of the OWASP Top 10 for LLM Applications?

The current edition is 2026, published by OWASP in August 2026. It replaces the 2025 edition, which was released in November 2024.

What changed between the 2025 and 2026 OWASP LLM lists?

The ten risk areas stayed and the order changed. Prompt Injection and Sensitive Information Disclosure remain first and second, Excessive Agency moved from sixth to third, Improper Output Handling moved from fifth to tenth, and System Prompt Leakage became Hidden Context Exposure.

Which OWASP LLM risks apply to an AI email assistant?

All ten have an email form, and the top three of the 2026 list matter most: Prompt Injection, Sensitive Information Disclosure and Excessive Agency. OWASP's own example for Excessive Agency is a mail assistant tricked by an incoming email into forwarding what it finds.

Does OWASP say prompt injection can be prevented?

No. The 2026 edition says no reliable prevention mechanism exists today. It advises designing the system on the assumption that the model will be fooled, and limiting what the model can do and what its output can reach.

Sources

Give your agent an inbox of its own. Free for two inboxes. No card.