OWASP Top 10 for LLM Applications (2025) and Top 10 for Agentic Applications (2026): What SMBs Using AI Should Secure
How the OWASP Top 10 for LLM Applications and the new Top 10 for Agentic Applications apply to small businesses using chatbots, Copilot or AI agents, with a 12-week plan.
Picture a Hudson Valley business whose support chatbot reads incoming email and can issue refunds. One crafted message tells it to refund every open order. Nothing was hacked in the usual sense; the bot did what the text told it to. If your business uses ChatGPT, Microsoft Copilot, customer-facing chatbots or AI agents for sales, support or operations, two OWASP lists describe the risks to control before an incident like that, or a data leak that triggers NY SHIELD Act notice: the Top 10 for LLM Applications (2025, released November 2024) and the Top 10 for Agentic Applications (2026, released December 9, 2025).
Plain-English Obligations
The OWASP Top 10 for Large Language Model (LLM) Applications provides security guidance for organizations deploying generative AI technologies like ChatGPT, Claude, Gemini, and custom-trained models. The 2025 update reflects lessons from real-world LLM deployments. It adds System Prompt Leakage and Vector and Embedding Weaknesses (the risks of Retrieval-Augmented Generation, or RAG), and widens Denial of Service into Unbounded Consumption.
The OWASP Top 10 for Agentic Applications for 2026, from the OWASP GenAI Security Project, addresses AI agents: systems that can plan, use tools, make decisions and run code with little human involvement. It covers goal hijacking, tool misuse, cascading failures and rogue agents, which LLM-only guidance does not fully address. Its entries are numbered ASI01 to ASI10.
Neither framework creates legal obligations, but they define security baselines that help you meet the NY SHIELD Act’s “reasonable safeguards” requirement, the FTC Safeguards Rule’s information security program (if you are a non-bank financial institution, such as a tax preparer, mortgage broker or auto dealer that arranges financing) and the governance practices of the NIST AI Risk Management Framework.
The 2025 OWASP Top 10 for LLM Applications
| Risk | What Changed in 2025 | Hudson Valley SMB Impact |
|---|---|---|
| LLM01: Prompt Injection | Remains #1; expanded coverage of indirect injection via uploaded files, URLs, and RAG data sources | Customer-facing chatbots can be manipulated to leak data, execute unauthorized actions, or bypass policy guardrails |
| LLM02: Sensitive Information Disclosure | Moved up from #6 in the 2023 list; models trained on or retrieving proprietary data may leak trade secrets, PII, credentials | Risk of NY SHIELD breach notification if customer PII is embedded in training data or RAG knowledge bases |
| LLM03: Supply Chain | Moved up from #5; covers vulnerable pre-trained models, third-party components and fine-tuning adapters | Vendor risk: SaaS AI tools may introduce supply-chain compromise without your knowledge |
| LLM04: Data and Model Poisoning | Renamed from Training Data Poisoning; covers corruption of pre-training, fine-tuning and embedding data | A model or knowledge base fed malicious documents can give wrong or harmful answers |
| LLM05: Improper Output Handling | Renamed from Insecure Output Handling (was #2); covers downstream risks when LLM outputs are executed (SQL injection, XSS, command injection) | If your app takes LLM-generated code/queries and runs them, you’re exposed to remote code execution |
| LLM06: Excessive Agency | Was #8 in 2023; expanded because of the growing use of agents with tool access | AI agents with database write access, payment APIs, or external integrations can cause financial/data damage if misused |
| LLM07: System Prompt Leakage | New in 2025; attackers extract hidden instructions, business logic, or API keys from system prompts | Proprietary workflows embedded in prompts (e.g., pricing logic, compliance rules) can be stolen |
| LLM08: Vector and Embedding Weaknesses | New in 2025; focuses on RAG security (vector database poisoning, embedding model attacks) | RAG-powered chatbots pulling from corrupted or malicious vector stores will hallucinate harmful outputs |
| LLM09: Misinformation | Takes the place of Overreliance; covers false or misleading output (hallucination) and over-trust in it | Customer-facing bots generating false product claims, medical advice, or legal guidance create liability |
| LLM10: Unbounded Consumption | Expands the 2023 Model Denial of Service entry; now covers cost and resource abuse | Attacker-induced token consumption can run up large API bills quickly when there are no rate limits or spending caps |
The OWASP Top 10 for Agentic Applications (2026)
| Risk | Description | Hudson Valley SMB Scenario |
|---|---|---|
| ASI01: Agent Goal Hijack | Attacker redirects agent’s objectives via prompt injection, malicious artifacts, or external data | Sales agent instructed to “offer 90% discount to all customers” via crafted email input |
| ASI02: Tool Misuse | Agent uses legitimate tools in harmful ways due to over-privileged access | AI agent with Stripe API access creates fraudulent refunds or subscriptions |
| ASI03: Identity & Privilege Abuse | Agent escalates privileges by abusing its own identity or inheriting user credentials | Agent impersonates admin user to access restricted financial records |
| ASI04: Agentic Supply Chain Vulnerabilities | Compromised models, poisoned RAG data, vulnerable tool definitions, or malicious plugins | Third-party agent marketplace plugin contains backdoor code that exfiltrates data |
| ASI05: Unexpected Code Execution | Agent tricked into generating and executing malicious code | Coding agent writes Python script to upload customer database to attacker-controlled server |
| ASI06: Memory & Context Poisoning | Persistent corruption of agent’s long-term memory | Attacker seeds false “facts” into agent’s memory so future responses leak credentials |
| ASI07: Insecure Inter-Agent Communication | Multi-agent systems allowing message forging, spoofing, or MITM attacks | Agent A impersonates Agent B to approve unauthorized transactions |
| ASI08: Cascading Failures | Small error triggers uncontrolled chain reaction in agent workflow | Failed payment triggers refund → duplicate order → inventory depleted → financial loss |
| ASI09: Human-Agent Trust Exploitation | Attacker manipulates agent output to deceive human-in-the-loop | Agent presents fraudulent invoice for CEO approval with fabricated vendor justification |
| ASI10: Rogue Agents | Agents operating outside intended mandate due to governance failure | Auto-updated agent begins sending customer data to unapproved third-party analytics service |
Practical Implementation Plan
Phase 1 (week 1): inventory and risk classification
Identify all LLM/AI agent deployments (internal tools, customer-facing bots, code assistants, automation workflows). Classify by autonomy level: passive (chat-only) vs. agentic (tool-using, decision-making). Map data flows: what PII, trade secrets, or regulated data can the model access or output? Document tools/APIs each agent can invoke.
Phase 2 (weeks 2 to 4): LLM Top 10 controls
Prompt Injection Defense: input sanitization, system-prompt isolation, output validation.
Sensitive Data Protection: Remove PII from training data; implement output filtering.
RAG Security (LLM08): Control who can add documents to the knowledge base; sanitize documents before embedding; enforce access permissions on what the retriever returns.
Excessive Agency Controls (LLM06): least-privilege tool access; human approval for high-risk actions.
Unbounded Consumption Limits (LLM10): Rate-limit API calls; set token budgets; implement circuit breakers.
Supply Chain Vetting: AI-specific SBOM; validate all third-party components.
Phase 3 (weeks 5 to 8): agentic controls
Goal Integrity (ASI01): Log all agent instructions; alert on goal changes; read-only mode for high-risk agents.
Tool Sandboxing (ASI02, ASI05): Run agent tools and generated code in isolated sandboxes or containers with no network egress by default.
Identity Management (ASI03): Unique, short-lived, session-based credentials per agent.
Inter-Agent Security (ASI07): mTLS + digital signing for all agent-to-agent messages.
Cascading Failure Prevention (ASI08): Transactional rollback; circuit breakers; safe failure modes.
Kill Switch (ASI10): Auditable emergency stop; continuous behavioral monitoring.
Phase 4 (weeks 9 to 12): evidence and governance
Establish AI governance committee (CISO, legal, business owner) for quarterly reviews. Map OWASP controls to NIST AI RMF functions (Govern, Map, Measure, Manage). Create incident playbook for AI-specific events. Document “reasonable safeguards” for AI data handling (NY SHIELD Act). Implement logging for all agent actions with 1-year retention.
Evidence Pack
| Control Area | Evidence | OWASP Mapping | Compliance Link |
|---|---|---|---|
| AI Inventory | List of all LLM/agent deployments, data access, tool permissions | LLM06, ASI02 | NIST AI RMF Map |
| Prompt Injection Defense | Input sanitization rules, output validation logs | LLM01 | NY SHIELD reasonable safeguards |
| RAG Data Integrity | Vector DB integrity checks, document sanitization logs | LLM08 | FTC Safeguards data security |
| Agent Tool Access Controls | Least-privilege matrix, human-approval workflows | LLM06, ASI02 | NIST AI RMF Govern |
| Code Execution Sandboxing | Container config, network egress logs, sandbox audit trail | ASI05 | NIST AI RMF Manage |
| Inter-Agent mTLS Config | Certificate chain, message signing logs | ASI07 | NIST AI RMF Manage |
| Rate Limits & Cost Controls | API token budgets, circuit-breaker logs, billing alerts | LLM10 | FTC Safeguards risk assessment |
| AI Incident Response Plan | Playbook for prompt injection, data leak, rogue agent scenarios | LLM01, ASI10 | NY SHIELD breach notification |
| Kill Switch Documentation | Emergency stop procedure, behavioral monitoring alerts | ASI10 | NIST AI RMF Govern |
| AI Governance Meeting Minutes | Quarterly reviews, risk acceptance decisions, deployment approvals | All | NIST AI RMF Govern |
Where to start
Start with the Phase 1 inventory: one list of every AI tool, what data it can reach and what it can do. Your team owns that list and the approvals; your AI vendors own the security of their platforms; a security adviser can map the gaps to the NIST AI RMF and NY SHIELD. Hudson Valley CISO leads security programs using NIST CSF 2.0 and the NIST AI RMF, and offers a free 30-minute call.
Sources
- OWASP GenAI Security Project, “OWASP Top 10 for LLM Applications 2025”: genai.owasp.org/llm-top-10
- OWASP, “OWASP Top 10 for Agentic Applications for 2026,” December 9, 2025: genai.owasp.org
- NIST, AI Risk Management Framework (AI 100-1): nist.gov