The OWASP Top 10 for LLM Applications currently ranks ten risk categories, per the 2026 edition the OWASP GenAI Security Project published on 4 August 2026, superseding the first widely cited 2025 edition. It is a community-developed awareness document that ranks the most critical security risks facing applications built on large language models. The project describes its output as written "for developers, architects, data scientists, security practitioners, and organizations building or operating applications that use large language models." It follows the same format OWASP has used since 2003 for its web application Top 10 (see our OWASP Top 10:2025 guide): a ranked, consensus-built list of risk categories rather than a scoring formula or compliance checklist.
Key takeaways
- The OWASP Top 10 for LLM Applications currently ranks ten risk categories, per the 2026 edition the OWASP GenAI Security Project published on 4 August 2026, superseding the first widely cited 2025 edition.
- Both, because sources still quote both.
- The 2026 Preface explains the moves.
- For a device maker, the useful reading is which of the ten risks the product's AI feature actually creates.
Which editions are worth knowing: 2025 or 2026?
Both, because sources still quote both. A 2025 edition, released in March 2025, was the first widely cited version. The GenAI Security Project has since published an OWASP GenAI LLM Top 10 2026 edition — the publication page states this release was published 4 August 2026. OWASP's original project repository for the 2025-era list (OWASP/www-project-top-10-for-large-language-model-applications) now carries a notice that it "is maintained as a legacy entry point and historical archive," with active development moved to a dedicated GenAI-Security-Project/GenAI-LLM-Top10 repository, whose 2026/final/ folder holds the canonical Markdown text. Anyone linking to or quoting "the OWASP LLM Top 10" should check which edition a source is describing, since the item names, codes and order differ between them.
What are the ten risks in the current (2026) edition?
Per the current GenAI-Security-Project repository:
| Code | Risk |
|---|---|
| LLM01:2026 | Prompt Injection |
| LLM02:2026 | Sensitive Information Disclosure |
| LLM03:2026 | Excessive Agency |
| LLM04:2026 | Supply Chain |
| LLM05:2026 | Data and Model Poisoning |
| LLM06:2026 | Unbounded Consumption |
| LLM07:2026 | Misinformation |
| LLM08:2026 | Hidden Context Exposure |
| LLM09:2026 | Vector and Embedding Weaknesses |
| LLM10:2026 | Improper Output Handling |
What moved between 2025 and 2026?
In the 2025 edition the order was Prompt Injection, Sensitive Information Disclosure, Supply Chain, Data and Model Poisoning, Improper Output Handling, Excessive Agency, System Prompt Leakage, Vector and Embedding Weaknesses, Misinformation, and Unbounded Consumption. Comparing the two lists: Prompt Injection and Sensitive Information Disclosure keep the top two positions; Excessive Agency and Supply Chain both move up; Unbounded Consumption rises four places; Improper Output Handling drops from fifth to tenth; and System Prompt Leakage gives way to Hidden Context Exposure as item eight.
The 2026 Preface explains the moves. It calls Excessive Agency's climb to third "the most consequential move on the list, because the vote and the record agree that agentic deployments are where the damage is landing." On item eight it says: "What used to be System Prompt Leakage is now Hidden Context Exposure, a broader framework for the same failure to trust information that should have stayed out of reach." The LLM08 text defines hidden context as more than the system prompt: developer instructions, retrieved policy text, tool and function schemas, and other material the application puts in the model's context.
Several entries also grew in scope, per the Preface. Prompt Injection now covers cross-modal attacks hidden "inside an image or an audio track". Supply Chain covers model artifacts that are not what they claim to be. Data and Model Poisoning absorbs fine-tuning subversion, and Improper Output Handling spans insecure code generated by assistants.
What does each risk look like in a connected product?
For a device maker, the useful reading is which of the ten risks the product's AI feature actually creates. Wiring determines most of the answer. A cloud chatbot on the support site, a voice assistant that can change device settings, and a small model running on the device each produce a different set. The table pairs each 2026 item with a device-side example and the first control the OWASP text gives for it.
| 2026 item | Example in a connected product | First control in the OWASP text |
|---|---|---|
| LLM01 Prompt Injection | A camera's vision-language feature reads instructions printed on a sign in frame | Filter at every modality boundary (text, image, audio), not just text |
| LLM02 Sensitive Information Disclosure | The assistant repeats another household member's schedule or a support ticket's contents | Authorize before retrieval, inside the index query; never store secrets in system prompts |
| LLM03 Excessive Agency | A "helpful" agent can call unlock_door or factory_reset |
Minimise tools, tool functions and tool permissions; require user approval for high-impact actions |
| LLM04 Supply Chain | A bundled open-weight model or LoRA adapter is swapped in the update pipeline | Signed, hash-pinned model artifacts and an SBOM extended to models (AIBOM) |
| LLM05 Data and Model Poisoning | Fine-tuning data scraped from user forums carries planted instructions | Track dataset and model lineage with an SBOM / ML-BOM; validate incoming data |
| LLM06 Unbounded Consumption | A scripted client drives the product's paid inference endpoint with long prompts | Token-aware cost controls and hard spending caps, not only request-rate limits |
| LLM07 Misinformation | The assistant invents a reset procedure or a safety threshold | Ground claims in authoritative, current sources before acting (claim-check-act) |
| LLM08 Hidden Context Exposure | The system prompt contains the cloud API key for device control | Keep secrets out of hidden context; enforce rules outside the model |
| LLM09 Vector and Embedding Weaknesses | One customer's manuals and logs are retrievable by another tenant | Permission and access control on the store; segregate data by trust tier |
| LLM10 Improper Output Handling | Model output is passed into a shell, a device command or a Markdown view in the mobile app | Treat the model as any other user; encode by context; disable auto-rendered images and link previews |
The ranks give a starting order, not a severity verdict for a given product. A read-only summariser of manuals has almost no LLM03 exposure, while an agent with device-control tools makes LLM03 and LLM01 the pair to design around. Our guide to prompt injection in AI-enabled IoT devices goes deeper on the first; the OWASP agentic AI threats guide covers multi-step agents, which the LLM08 text explicitly leaves to the separate OWASP Top 10 for Agentic Applications.
What does the 2026 list say about LLMs that ship on the device?
It treats device and firmware integrity as part of the model's supply chain. LLM04 has a section titled "On-Device LLM Supply-Chain Vulnerabilities," which states that "LLMs shipped on devices add compromised manufacturing processes, exploitation of device OS or firmware vulnerabilities, and re-packaged applications with tampered models to the attack surface, making device integrity and firmware trust part of the LLM supply chain."
The mitigations listed in the same entry read like a firmware team's checklist:
- maintain a signed inventory of components using an SBOM, "extended to models, adapters, and datasets with AI BOMs (AIBOMs) and ML SBOMs," with OWASP CycloneDX ML-BOM named as one option;
- use models only from verifiable sources, with signing and file hashes, naming the OpenSSF Model Signing project and Sigstore;
- "Encrypt models deployed at the edge with integrity checks, use vendor attestation APIs to prevent tampered apps and models, and reject unrecognized firmware and untrusted device states."
LLM04 also warns that unsafe serialization formats such as Python pickle "can execute arbitrary code on load." In practice the model file belongs in the same secure-boot and signed-update chain as the firmware, and the model runtime libraries belong in the firmware SBOM. Our AI-BOM guide for embedded AI devices and the firmware SBOM pillar cover the inventory side.
Which controls have to live outside the model?
Anything that decides whether an action is allowed. The 2026 text is consistent on this across entries: the model is a component that can be steered, so authorization, secrets and policy enforcement sit in deterministic code around it. Drawn from LLM01, LLM03, LLM08 and LLM10:
Hold credentials in application code, not in the promptLLM08: "Assume all context available to the LLM could also be available to users." LLM01 asks for credentials and state-change capability to stay "in application code, not the model." Execute tools in the user's contextLLM03 asks that actions run on downstream systems "in the context of that specific user, and with the minimum privileges necessary." For a device cloud, that means the agent's call to unlock a door goes through the same device-binding check as the app's call (see our OWASP API Security Top 10 guide). Apply complete mediationLLM03: "Implement authorization in logic rather than relying on an LLM to decide if an action is allowed or not." It suggests graduated enforcement (audit, warn, block, escalate) so reversible actions can auto-approve and irreversible ones go to a human. Budget capabilities with the Rule of TwoLLM01 treats simultaneous access to untrusted input, sensitive data, and state change or external communication as high-risk, needing per-action human approval when all three are present. Validate output before anything acts on itLLM10: "Treat the model as any other user," validate structured output in trusted code, and in mobile and chat clients "Disable auto-rendering of Markdown images, link previews, iframes, and similar elements by default."
Worked example: how would you review a voice assistant that controls a smart lock?
Synthetic example. The product and findings are invented for illustration and do not describe any real vendor.
A smart-lock maker adds a cloud voice assistant to its app. The model can call four tools: get_lock_state, lock, unlock and create_guest_code. The team reviews it against the 2026 list:
Map the Rule of Two (LLM01)The assistant takes untrusted input (voice, plus calendar text it reads for guest arrivals), holds sensitive data (guest names), and changes state ( unlock). All three are present, sounlockandcreate_guest_codeneed per-action confirmation in the app, showing the exact action.Check hidden context (LLM08)The system prompt holds a service token for the lock API. They move the token into the tool service and pass only the user's own session through. Check tool scope (LLM03) create_guest_codeaccepts any lock serial. They make the tool resolve the serial from the user's bindings rather than from model output.Check output handling (LLM10)Replies render as Markdown in the app, including remote images. They switch off image auto-loading. Check cost (LLM06)No per-account token cap exists. They add a daily cap and alerting. Check supply chain (LLM04)The speech model on the hub ships unsigned in the update bundle. They add it to the signed image and to the SBOM.
What are the common pitfalls?
- Treating the system prompt as a security boundary. LLM08 says hidden context should not "be solely relied upon as a security boundary for authorization, privilege separation, policy enforcement, or content filtering."
- Filtering text only. A camera or microphone is an input channel; LLM01 names image, audio and video injection.
- Letting the agent use a service account. A broad service identity turns any successful injection into fleet-wide control.
- Leaving the model file out of the update chain. An unsigned model in an otherwise signed image is the on-device case LLM04 describes.
- Mixing up item codes across editions. LLM07 meant System Prompt Leakage in 2025 and means Misinformation in 2026.
Why should device and firmware makers care?
This list targets LLM-application risks specifically — prompt handling, model supply chain, output validation, agentic tool access — not the conventional application or network risks a product may also carry. A camera with an LLM-driven voice assistant still has firmware, network and authentication attack surface covered by the OWASP IoT Top 10 and the OWASP Top 10:2025 web risks, plus whichever of these ten risks apply to how the model is prompted, what data it can see, and what actions it may take.
SoftWSP does not red-team or test LLM features. Its part is the firmware those features run on. It extracts a software bill of materials from a firmware image as CycloneDX 1.6 JSON, which includes model-runtime and inference libraries if they ship in the image. It matches the components against NVD, the CISA KEV catalogue, FIRST EPSS scores and CISA ICS advisories, and drafts a cited report for an engineer to review. Try the free SBOM viewer, or see device firmware security for scoped work.
This guide summarises public OWASP GenAI Security Project resources as fetched in November 2026; it does not constitute legal or compliance advice.
Put this guide to work
Frequently asked questions
What is the OWASP Top 10 for LLM Applications?
A community-developed awareness document, published by the OWASP GenAI Security Project, that ranks the most critical security risks in applications built on large language models. The project's own description states it is written "for developers, architects, data scientists, security practitioners, and organizations building or operating applications that use large language models," the same awareness-document format OWASP uses for its long-running web and API Top 10 lists (see our [OWASP Top 10:2025 guide](/guides/owasp-top-10-2025/)).
What are the ten risks in the 2026 edition?
Per the GenAI Security Project's GitHub repository for the 2026 release: LLM01:2026 Prompt Injection, LLM02:2026 Sensitive Information Disclosure, LLM03:2026 Excessive Agency, LLM04:2026 Supply Chain, LLM05:2026 Data and Model Poisoning, LLM06:2026 Unbounded Consumption, LLM07:2026 Misinformation, LLM08:2026 Hidden Context Exposure, LLM09:2026 Vector and Embedding Weaknesses, and LLM10:2026 Improper Output Handling.
What changed from the 2025 list to the 2026 list?
The ranking order moved and one named item changed. In the 2025 edition the ten items were (in order) Prompt Injection, Sensitive Information Disclosure, Supply Chain, Data and Model Poisoning, Improper Output Handling, Excessive Agency, System Prompt Leakage, Vector and Embedding Weaknesses, Misinformation, and Unbounded Consumption. The 2026 edition reorders several of these (for example Excessive Agency and Supply Chain move up, Improper Output Handling moves down to tenth) and replaces System Prompt Leakage with a newly named item, Hidden Context Exposure. The project's own GenAI-LLM-Top10 README states the new edition "combines community judgment with analysis of real-world incidents and updates the ordering, scope, examples, mitigations, and framework mappings across the list," without itemising every change line by line — organisations tracking a specific 2025 item should confirm where it landed in 2026 rather than assuming the code number carried over.
How does this relate to OWASP's other Top 10 lists?
It is a separate, purpose-built list for the LLM/generative-AI risk surface, not a replacement for OWASP's existing lists. A device or service that embeds an LLM feature can still have conventional web, API or IoT weaknesses — see our guides to the [OWASP Top 10:2025](/guides/owasp-top-10-2025/), [OWASP API Security Top 10](/guides/owasp-api-security-top-10/) and [OWASP IoT Top 10](/guides/owasp-iot-top-10/) — on top of whichever LLM-specific risks apply to how the model is integrated, prompted and given access to tools or data.
Sources
- OWASP GenAI Security Project: LLM Top 10 resource page
- OWASP GenAI Security Project: OWASP GenAI LLM Top 10 2026 (publication page)
- GitHub: GenAI-Security-Project/GenAI-LLM-Top10 (current repository, 2026 edition README)
- GitHub: OWASP/www-project-top-10-for-large-language-model-applications (legacy archive repository)
- OWASP GenAI LLM Top 10 2026 — Preface: What's New in the 2026 Top 10
- OWASP GenAI LLM Top 10 2026 — LLM01:2026 Prompt Injection
- OWASP GenAI LLM Top 10 2026 — LLM03:2026 Excessive Agency
- OWASP GenAI LLM Top 10 2026 — LLM04:2026 Supply Chain
- OWASP GenAI LLM Top 10 2026 — LLM06:2026 Unbounded Consumption
- OWASP GenAI LLM Top 10 2026 — LLM08:2026 Hidden Context Exposure
- OWASP GenAI LLM Top 10 2026 — LLM10:2026 Improper Output Handling
About this guide
SoftWSP Firmware Security Desk
Covers the Firmware vulnerabilities & fixes cluster: 129 guides.

In this cluster
Questions about this page?
