{"data":{"id":"e69b09e7-ccd3-42c7-8670-8e7293d3ae48","title":"10 most critical LLM vulnerabilities","summary":"Large language models can leak sensitive information or be manipulated through prompt injection (tricking an AI by hiding malicious instructions in its input) and other vulnerabilities that pose security, legal, and compliance risks to enterprises. OWASP has updated its list of the top 10 LLM vulnerabilities based on real-world incidents and expert analysis, with prompt injection and sensitive information disclosure remaining the most severe threats, while excessive agency (agentic actions outside permitted bounds) has risen in prominence as AI systems become more autonomous.","solution":"To reduce prompt injection risk, OWASP recommends: (1) constrain the model's role and capabilities in the system prompt, (2) add a human in the loop for sensitive operations requiring extra approval steps, (3) define a strict output schema and validate all responses with trusted application code, (4) scan inputs and outputs (text, image, audio, and structured data) for harmful content and block sensitive or unauthorized content before it reaches the model or is returned to users, and (5) hold credentials and state-change capability in application code rather than the model, granting least privilege per operation.","labels":["security","safety"],"sourceUrl":"https://www.csoonline.com/article/575497/owasp-lists-10-most-critical-large-language-model-vulnerabilities.html","publishedAt":"2026-09-10T08:25:00.000Z","cveId":null,"cweIds":null,"cvssScore":null,"cvssSeverity":null,"severity":"info","attackType":["prompt_injection","data_extraction","jailbreak"],"issueType":"news","affectedPackages":null,"affectedVendors":[],"affectedVendorsRaw":["OWASP"],"classifierModel":"claude-haiku-4-5-20251001","classifierPromptVersion":"v3","cvssVector":null,"attackVector":null,"attackComplexity":null,"privilegesRequired":null,"userInteraction":null,"exploitMaturity":null,"epssScore":null,"patchAvailable":null,"disclosureDate":"2026-09-10T08:25:00.000Z","capecIds":null,"crossRefCount":0,"attackSophistication":"moderate","impactType":["confidentiality","integrity","safety"],"aiComponentTargeted":null,"llmSpecific":true,"classifierConfidence":0.85,"researchCategory":null,"atlasIds":null}}