OWASP LLM Top 10
The OWASP Top 10 for Large Language Model Applications enumerates the ten most critical risks specific to LLM-integrated systems. Gadriel Code maps GVL findings directly onto LLM01–LLM10.
Related: Compliance overview · OWASP Agentic Top 10 · OWASP Agentic Skills Top 10.
At a glance
| Property | Value |
|---|---|
| Slug | owasp-llm-top10 |
| Aliases | owasp-llm, llm-top10 |
| Controls | 10 |
| Manual-only | 0 |
| Machine-assessable | 10 |
| Direct mapping key | owasp_llm |
KEY TAKEAWAY
Alongside the EU AI Act, this is the most automated framework: every one of LLM01–LLM10 is predicated and driven from source evidence.
Generating the report
bashgadriel code report --compliance owasp-llm-top10
Outputs: .security/compliance/owasp-llm-top10.{md,typ,pdf}.
What Gadriel evaluates
Each control corresponds to an LLM Top 10 category. Representative examples:
- LLM01 Prompt Injection — untrusted-input flows into prompt construction.
- LLM02 Sensitive Information Disclosure — secrets/PII in prompts or logs.
- LLM03 Supply Chain — model/plugin/dependency provenance and integrity.
- LLM04 Data & Model Poisoning — unvetted training/data pipelines.
- LLM05 Improper Output Handling — unsanitized model output flowing into downstream sinks.
- LLM06 Excessive Agency — over-permissioned tool calls.
- LLM07 System Prompt Leakage — system prompts exposed to users or logs.
- LLM08 Vector & Embedding Weaknesses — poisoning, retrieval-side auth bypass.
- LLM09 Misinformation — over-reliance patterns, missing user disclosures.
- LLM10 Unbounded Consumption — missing rate limits and cost controls.
Rules carry compliance_mapping.owasp_llm: LLM01 (etc.).