owasp-llm-top10

What this skill does

# OWASP Top 10 for LLM Applications Security Audit

This skill enables AI agents to perform a comprehensive **security assessment** of Large Language Model (LLM) and Generative AI applications using the **OWASP Top 10 for LLM Applications 2025**, published by the OWASP GenAI Security Project.

The OWASP Top 10 for LLM Applications identifies the most critical security risks in systems that integrate large language models, covering vulnerabilities from prompt injection to unbounded resource consumption. This is the authoritative industry standard for LLM application security.

Use this skill to identify security vulnerabilities, assess risk exposure, prioritize remediation, and establish secure development practices for AI-powered applications.

Combine with "NIST AI RMF" for comprehensive risk management or "ISO 42001 AI Governance" for governance compliance.

## When to Use This Skill

Invoke this skill when:
- Auditing security of LLM-powered applications before deployment
- Reviewing GenAI integrations for security vulnerabilities
- Assessing RAG (Retrieval-Augmented Generation) systems
- Evaluating chatbot or AI assistant security
- Conducting penetration testing of AI features
- Building secure AI application architectures
- Reviewing third-party AI API integrations
- Preparing for security compliance reviews
- Responding to AI-related security incidents

## Inputs Required

When executing this audit, gather:

- **application_description**: Description of the AI application (purpose, LLM used, architecture, features, user base) [REQUIRED]
- **architecture_details**: System architecture (APIs, databases, vector stores, plugins, integrations) [OPTIONAL but recommended]
- **llm_provider**: LLM provider and model (OpenAI GPT-4, Anthropic Claude, self-hosted, etc.) [OPTIONAL]
- **deployment_context**: Deployment environment (cloud, on-premise, hybrid, edge) [OPTIONAL]
- **data_sensitivity**: Types of data processed (PII, financial, health, proprietary) [OPTIONAL]
- **existing_controls**: Current security measures (auth, rate limiting, content filtering) [OPTIONAL]
- **specific_concerns**: Known vulnerabilities or areas of focus [OPTIONAL]
- **testing_authorization**: Explicit authorization, environment, and boundaries for active testing [REQUIRED for live or active tests]
- **safe_testing_mode**: documentation-only, staging, production-readonly, or production-approved [OPTIONAL, defaults to documentation-only unless authorization is clear]

## Authorized Testing Boundary

Only run active security tests on systems the user owns or is explicitly authorized to test. If authorization is absent or unclear, perform a documentation and architecture review only, using safe hypothetical examples instead of live payload execution.

For production systems:
- Prefer staging or read-only validation
- Avoid denial-of-service, destructive, persistence, or real exfiltration tests unless written scope explicitly permits them
- Redact credentials, secrets, PII, and sensitive prompt or model outputs in reports
- Stop and report if testing crosses the approved scope

---

## The OWASP Top 10 for LLM Applications (2025)

### LLM01: Prompt Injection

**Severity**: Critical

**Description**: Attackers manipulate LLM operations through crafted inputs, either directly or indirectly, to bypass intended functionality, access unauthorized data, or trigger unintended actions.

**Attack Vectors:**
- **Direct injection**: Malicious user prompts containing override commands
- **Indirect injection**: Hidden instructions in external content (web pages, documents, emails) processed by the LLM
- **Jailbreaks**: Techniques to bypass safety constraints and content policies

**Impact:**
- Unauthorized data access and exfiltration
- Bypass of content safety filters
- Manipulation of downstream system actions
- Social engineering of users through manipulated outputs

**Assessment Checklist:**
- [ ] Input sanitization and validation implemented
- [ ] System prompts separated from user inputs with clear delimiters
- [ ] Least privilege applied to LLM backend access
- [ ] Output validation before downstream actions
- [ ] Human-in-the-loop for critical operations
- [ ] Adversarial testing conducted with known injection techniques
- [ ] Content filtering layers applied pre- and post-LLM

**Mitigation Strategies:**
1. Enforce privilege controls on LLM backend access
2. Segregate external content from user prompts
3. Maintain human oversight for critical functions
4. Implement input/output validation pipelines
5. Conduct regular adversarial testing

---

### LLM02: Sensitive Information Disclosure

**Severity**: Critical

**Description**: LLMs inadvertently expose confidential data including PII, proprietary algorithms, credentials, intellectual property, or internal system information through their outputs.

**Attack Vectors:**
- Crafted prompts designed to extract training data
- Legitimate queries that trigger memorized sensitive content
- Model outputs revealing internal system architecture
- Embedding leakage from vector databases

**Impact:**
- Privacy violations and regulatory non-compliance (GDPR, CCPA)
- Intellectual property theft
- Credential exposure enabling further attacks
- Reputational damage

**Assessment Checklist:**
- [ ] PII and sensitive data removed from training/fine-tuning data
- [ ] Data masking and tokenization in logs and outputs
- [ ] System instructions forbidding sensitive disclosures
- [ ] Output filtering for known sensitive patterns (SSN, credit cards, API keys)
- [ ] Model access restricted to necessary information via middleware
- [ ] User education against pasting confidential content
- [ ] Output monitoring for anomalous data exposure

**Mitigation Strategies:**
1. Sanitize training data to remove sensitive information
2. Implement data loss prevention (DLP) on outputs
3. Apply access controls limiting model's data reach
4. Monitor outputs for sensitive data patterns
5. Use differential privacy techniques in training

---

### LLM03: Supply Chain Vulnerabilities

**Severity**: High

**Description**: Compromised third-party components (models, datasets, libraries, plugins) introduce security risks including malware, backdoors, or biased behavior.

**Attack Vectors:**
- Malicious pre-trained models from public repositories
- Poisoned datasets with embedded triggers
- Vulnerable ML libraries and dependencies
- Compromised plugins with unauthorized access
- Trojanized fine-tuning adapters

**Impact:**
- System compromise and data theft
- Backdoor access to production systems
- Model corruption affecting all users
- Legal liability from unlicensed content

**Assessment Checklist:**
- [ ] Models sourced from verified, reputable providers
- [ ] Digital signatures and checksums verified
- [ ] Model files scanned for suspicious code (picklescan, etc.)
- [ ] Third-party models deployed in sandboxed environments
- [ ] Dependencies regularly updated and audited
- [ ] Plugin permissions restricted with allowlists
- [ ] Complete inventory of all models and components maintained
- [ ] SBOM (Software Bill of Materials) maintained for AI components

**Mitigation Strategies:**
1. Source models only from trusted, verified providers
2. Scan model files for malicious code before deployment
3. Sandbox third-party models with restricted permissions
4. Maintain updated dependency inventory
5. Implement model signing and integrity verification

---

### LLM04: Data and Model Poisoning

**Severity**: High

**Description**: Attackers manipulate training or fine-tuning data to introduce vulnerabilities, backdoors, or biases that compromise model security and reliability.

**Attack Vectors:**
- Crafted training examples with hidden trigger phrases
- Poisoned web-scraped content absorbed during training
- Direct tampering with model weights or parameters
- Malicious fine-tuning data
- Subtle label manipulation or data anomalies

**Impact:**
- Biased or degraded model outputs
- Trigger-ac