Prompt Injection Remains Top Threat in LLM Security, OWASP Reports

The OWASP released its third version of the Top 10 for LLM Applications on August 4, 2026, ranking prompt injection as the leading security threat for the third consecutive year. Despite a low number of recorded incidents, security practitioners emphasize the significant risks associated with prompt injection, where user input can manipulate an LLM's behavior, potentially leading to harmful content or sensitive data disclosure. The report highlights that sensitive information disclosure is the second-highest threat, followed by excessive agency and misinformation. OWASP recommends designing systems with the assumption that instruction boundaries will be bypassed to mitigate these risks. The report indicates that while prompt injection remains a critical concern, actual incidents have been relatively few. The analysis reflects ongoing efforts in the cybersecurity community to address these vulnerabilities in LLM applications.

Key Points: • Prompt injection is ranked as the top threat to LLM applications for three years running. • Sensitive information disclosure and excessive agency are also significant risks. • OWASP advises designing systems to assume instruction boundaries will be bypassed.