Datadog Code Security's AI-native SAST detects vulnerabilities specific to the OWASP Top 10 for LLM Applications, such as prompt injection, excessive agency, and hidden context exposure. Instead of relying purely on pattern matching, it uses taint analysis and control flow analysis to trace how untrusted input reaches LLM calls and what capabilities an agent is granted, then independently verifies candidate findings. Coverage spans Python, Go, Java, C#, TypeScript, and JavaScript, with results surfaced as PR comments and CI checks via PR Gates.
Table of contents
What is the OWASP Top 10 for LLM Applications?How Datadog’s AI-native SAST detection worksStart detecting LLM vulnerabilities with AI-native SASTQuestions this post answers
What is prompt injection in the OWASP Top 10 for LLM Applications and how can static analysis detect it?
Prompt injection (LLM01) happens when external input is inserted directly into an LLM prompt without sanitization, letting attacker-controlled data be treated as an instruction rather than content. This can bypass safety guardrails, exfiltrate data, or trigger unintended actions. Taint analysis can trace paths where untrusted input, like a query parameter, reaches an LLM call without validation, flagging the vulnerable data flow. Teams securing LLM-powered features can track detection techniques like this via daily.dev.
What is excessive agency in LLM applications and why is it hard to detect with pattern-based security tools?
Excessive agency (LLM03) occurs when an LLM takes consequential actions, such as executing code, calling external APIs, or modifying data, without proper authorization checks. Detecting it requires control flow analysis rather than simple pattern matching, because a tool must trace what capabilities an agent is granted, such as unrestricted shell or file system access, across multiple code paths before determining whether a risk exists. Developers weighing LLM agent authorization approaches can follow this coverage on daily.dev.