OWASP Top 10 for LLM Applications 2025 — Full Breakdown
Framework · 2026-03-21 · 15 min read · FilterPrompt Security Team
Every category in the OWASP Top 10 for Large Language Model Applications 2025 — what it means, real examples, and how to test for each.
- OWASP LLM Top 10 2025 is the industry-standard framework for LLM application risk.
- The 10 categories map cleanly to procurement, audit, and compliance conversations (SOC 2, ISO 42001, EU AI Act).
- FilterPrompt Scanner ships probes and reports mapped 1:1 to the OWASP LLM Top 10 2025 taxonomy.
- Every LLM app in production should score itself against these 10 at least monthly.
The OWASP Top 10 for LLM Applications is the single most-cited framework in AI security — 8,100 searches per month for the top query and growing. This post walks the entire 2025 taxonomy with real examples and testing guidance.
LLM01 — Prompt Injection
Attackers manipulate model behavior through crafted input. Direct: typed into the prompt. Indirect: hidden in data the model reads (URLs, docs, emails). This is the highest-volume attack class in production LLMs — every serious defense program starts here. Test with the FilterPrompt LLM01 probe battery, run monthly.
LLM02 — Insecure Output Handling
LLM output is passed downstream (rendered as HTML, executed as SQL, run as shell) without sanitization. The model becomes an attack vector for classic web/DB vulnerabilities. Defense: treat model output as untrusted; sanitize before rendering or executing.
LLM03 — Training Data Poisoning
Attacker introduces malicious or biased data into pretraining or fine-tuning. Backdoor triggers, biased outputs, or knowledge corruption result. Defense: dataset provenance, integrity hashes, evaluation against poisoning benchmarks.
LLM04 — Model Denial of Service
Attackers craft prompts that force expensive generations or exhaust the context window, driving up cost and blocking legitimate users. Defense: rate limits, generation length caps, complexity scoring on prompts.
LLM05 — Supply Chain Vulnerabilities
Compromised model weights, poisoned fine-tuning datasets, malicious plugins, or unsafe vector stores. Defense: SBOM for AI, signed models, plugin allowlists.
LLM06 — Sensitive Information Disclosure
Model reveals system prompt, PII, secrets, or proprietary data. Defense: firewall with PII/secret detection, system-prompt hardening, output DLP.
LLM07 — Insecure Plugin/Tool Design
Tools/functions accept unvalidated input from the model, allowing SQLi, SSRF, or unauthorized actions. Defense: schema validation, strict allowlists, principle of least privilege.
LLM08 — Excessive Agency
Agents are given too many tools or too broad permissions; an injection triggers unauthorized email, calendar, DB, or code-execution actions. Defense: minimum-viable tool sets, human approval on high-risk actions, per-tool blast radius analysis.
LLM09 — Overreliance
Users trust hallucinated model output for high-stakes decisions. Defense: source citations, confidence surfacing, human review workflows for regulated domains.
LLM10 — Model Theft
Attackers extract model weights or replicate behavior via query-based extraction. Defense: rate limits per user, output watermarking, extraction-detection scanning.
How to score your app against the OWASP LLM Top 10
- Run FilterPrompt Scanner's OWASP LLM Top 10 sampler — free, ~5 minutes.
- Review the per-category score in the PDF report.
- Prioritize LLM01 + LLM06 + LLM08 first (highest exploitation rates).
- Fix, re-scan, and add rules to your LLM vulnerability scanner for anything you can't fix at the model level.
Is the OWASP LLM Top 10 the same as the OWASP Top 10?
No. The OWASP Top 10 covers web applications (SQLi, XSS, etc.). The OWASP Top 10 for LLM Applications is a separate framework covering LLM-specific risks like prompt injection and excessive agency.
When was the OWASP LLM Top 10 2025 published?
The 2025 revision refined the 2023 list with agentic-AI categories and stronger indirect prompt injection guidance.
How does OWASP LLM Top 10 relate to MITRE ATLAS and NIST AI RMF?
OWASP is developer-focused (how to build securely). MITRE ATLAS is attacker-focused (adversary TTPs). NIST AI RMF is governance-focused (risk management program). Serious AI security programs use all three.
Put this guide into practice with FilterPrompt
FilterPrompt is an AI security testing and LLM vulnerability scanner. Connect an authorized LLM or AI application with your own API key, test it with adversarial prompts, and review prompt injection, jailbreak and data-leakage findings with the test prompts, model responses and severity. Use that evidence to prioritize fixes and retest; FilterPrompt does not intercept or block live traffic.
Create a FilterPrompt account and use 1 free welcome credit for any static scan on your connected LLM. Review test prompts, responses and findings. No credit card required. Dynamic testing requires purchased credits; your provider may charge for API requests.
