Index of every page on this site: llms.txt
All skills
semgrep avatar

/llm-security

@68177b8 official
by semgrepsemgrep/skills350 stars
36
Run it yourself

Security guidelines for LLM applications based on OWASP Top 10 for LLM 2025. Use when building LLM apps, reviewing AI security, implementing RAG systems, or asking about LLM vulnerabilities like 'prompt injection' or 'check LLM security'. IMPORTANT: Always consult this skill when building chatbots, AI agents, RAG pipelines, tool-using LLMs, agentic systems, or any application that calls an LLM API (OpenAI, Anthropic, Gemini, etc.) — even if the user doesn't explicitly mention security. Also use when users import 'openai', 'anthropic', 'langchain', 'llamaindex', or similar LLM libraries.

Use this Skill: https://skilld.dev/gh/semgrep/skills/llm-security

Nothing lands on disk. Nothing to clean up.

Fork this Skill

Edit a local copy. It keeps the author and licence.

SKILL.md

≈152 tokens for metadata: the name and description. ≈1.1k when used: this file. ≈55k more on demand in 13 files.

Description uses 7.6% of example budget

Before choosing a Skill, your Agent reads its name and description. All available Skills share that space.

  • A shorter description leaves more room for other Skills. This entry exceeds our 1% size suggestion.

In our Claude Code example, all Skill names and descriptions share 8,000 characters. This Skill uses ≈608 characters, or 7.6%.

The 1% threshold is a size suggestion. Longer descriptions can still fit.

Your model, settings, and other Skills decide how much text your Agent can read.

Example settings and source

The example uses a 200k-token context and default Claude Code settings. The count includes the name, description, separators, and when_to_use when present. Codex also counts local file paths.

Skit's source and limits: Codex 0.160.1, Claude Code 2.1.292.

LLM Security Guidelines (OWASP Top 10 for LLM 2025)

Security rules for building secure LLM applications, based on the OWASP Top 10 for LLM Applications 2025.

How to Use This Skill

Proactive mode — When building or reviewing LLM applications, automatically check for relevant security risks based on the application pattern. You don't need to wait for the user to ask about LLM security.

Reactive mode — When the user asks about LLM security, use the mapping below to find relevant rule files with detailed vulnerable/secure code examples.

Workflow

  1. Identify what the user is building (see "What Are You Building?" below)
  2. Check the priority rules for that pattern
  3. Read the specific rule files from rules/ for code examples
  4. Apply the secure patterns or flag vulnerable ones

What Are You Building?

Use this to quickly identify which rules matter most for the user's task:

Building... Priority Rules
Chatbot / conversational AI Prompt Injection (LLM01), System Prompt Leakage (LLM07), Output Handling (LLM05), Unbounded Consumption (LLM10)
RAG system Vector/Embedding Weaknesses (LLM08), Prompt Injection (LLM01), Sensitive Disclosure (LLM02), Misinformation (LLM09)
AI agent with tools Excessive Agency (LLM06), Prompt Injection (LLM01), Output Handling (LLM05), Sensitive Disclosure (LLM02)
Fine-tuning / training Data Poisoning (LLM04), Supply Chain (LLM03), Sensitive Disclosure (LLM02)
LLM-powered API Unbounded Consumption (LLM10), Prompt Injection (LLM01), Output Handling (LLM05), Sensitive Disclosure (LLM02)
Content generation Misinformation (LLM09), Output Handling (LLM05), Prompt Injection (LLM01)

Categories

Critical Impact

  • LLM01: Prompt Injection (rules/prompt-injection.md) - Prevent direct and indirect prompt manipulation
  • LLM02: Sensitive Information Disclosure (rules/sensitive-disclosure.md) - Protect PII, credentials, and proprietary data
  • LLM03: Supply Chain (rules/supply-chain.md) - Secure model sources, training data, and dependencies
  • LLM04: Data and Model Poisoning (rules/data-poisoning.md) - Prevent training data manipulation and backdoors
  • LLM05: Improper Output Handling (rules/output-handling.md) - Sanitize LLM outputs before downstream use

High Impact

  • LLM06: Excessive Agency (rules/excessive-agency.md) - Limit LLM permissions, functionality, and autonomy
  • LLM07: System Prompt Leakage (rules/system-prompt-leakage.md) - Protect system prompts from disclosure
  • LLM08: Vector and Embedding Weaknesses (rules/vector-embedding.md) - Secure RAG systems and embeddings
  • LLM09: Misinformation (rules/misinformation.md) - Mitigate hallucinations and false outputs
  • LLM10: Unbounded Consumption (rules/unbounded-consumption.md) - Prevent DoS, cost attacks, and model theft

See rules/_sections.md for the full index with OWASP/MITRE references.

Quick Reference

Vulnerability Key Prevention
Prompt Injection Input validation, output filtering, privilege separation
Sensitive Disclosure Data sanitization, access controls, encryption
Supply Chain Verify models, SBOM, trusted sources only
Data Poisoning Data validation, anomaly detection, sandboxing
Output Handling Treat LLM as untrusted, encode outputs, parameterize queries
Excessive Agency Least privilege, human-in-the-loop, minimize extensions
System Prompt Leakage No secrets in prompts, external guardrails
Vector/Embedding Access controls, data validation, monitoring
Misinformation RAG, fine-tuning, human oversight, cross-verification
Unbounded Consumption Rate limiting, input validation, resource monitoring

Key Principles

  1. Never trust LLM output - Validate and sanitize all outputs before use
  2. Least privilege - Grant minimum necessary permissions to LLM systems
  3. Defense in depth - Layer multiple security controls
  4. Human oversight - Require approval for high-impact actions
  5. Monitor and log - Track all LLM interactions for anomaly detection

References

Source: SKILL.md on GitHub

No rule matched.

skilld matched fixed text patterns in SKILL.md and file names. Patterns miss obfuscated code.

skilld run checks every file with the same patterns. It asks for approval before it loads a Skill with a behavior marked Needs approval.

1 alert2mo5 checks · Risk SAFE
  • Gen Agent Trust Hub2mo

    The skill provides comprehensive security guidelines and code examples for building secure LLM applications, based on the OWASP Top 10 for LLMs 2025. It serves as an educational resource to help developers mitigate risks like prompt injection and sensitive data exposure. While the skill contains examples of vulnerable code with hardcoded credentials, these are clearly labeled as insecure patterns for demonstration and educational purposes, using non-functional example values.

  • Socket2mo

    No alerts

  • Snyk2mo

    Risk: MEDIUM · 1 issue

  • Runlayer8mo

    12/14 files flagged

  • ZeroLeaks7mo

    Score: 93/100 · 2 sections analyzed

Signed by skilld at 68177b8. This ties the file your Agent reads to that commit on GitHub. It does not review the instructions.

Last checked against GitHub last month.

Activeupdated 9 months ago
  • llm-security
  • prompt-injection
  • owasp
  • rag
  • ai-agents
  • data-poisoning
  • sensitive-disclosure
  • output-handling
  • vector-embeddings

README badge

README badge for semgrep/skills/llm-security

Provides security guidelines for LLM applications based on OWASP Top 10 for LLM 2025, covering prompt injection, sensitive disclosure, data poisoning, excessive agency, and other LLM-specific risks. Use when building chatbots, RAG systems, AI agents, or any application calling LLM APIs to identify relevant vulnerabilities and apply secure patterns.

Generated from the current SKILL.md.

Does this skill cover prompt injection attacks?
Yes. Prompt Injection (LLM01) is a critical category with dedicated rules for preventing both direct and indirect prompt manipulation in chatbots, RAG systems, and tool-using LLMs.
What LLM APIs and libraries does this apply to?
This skill applies to any application calling OpenAI, Anthropic, Gemini, or similar LLM APIs, and to code using LangChain, LlamaIndex, or comparable LLM frameworks.
Should I use this skill only when the user explicitly asks about security?
No. The skill is designed for proactive use: automatically check for relevant security risks whenever building or reviewing LLM applications, chatbots, RAG pipelines, or AI agents — regardless of whether the user mentions security.
Does this cover RAG system security?
Yes. RAG systems have priority rules for Vector/Embedding Weaknesses (LLM08), Prompt Injection (LLM01), Sensitive Disclosure (LLM02), and Misinformation (LLM09).
Are there code examples for each vulnerability?
Yes. Each of the 10 OWASP categories has a dedicated rule file in `rules/` with vulnerable and secure code examples.

Generated from the current SKILL.md. These answers refresh after source changes.