Find LLM freeform output flowing into a structured parser (OWASP LLM01)
org.openrewrite.ai.security.FindMissingStructuredOutputArtifactorg.openrewrite.recipe:rewrite-java-securityFind code where the plain-text output of an LLM chat call flows into a structured parser (ObjectMapper.readValue/readTree, Gson.fromJson). Model output is read through the LangChain4j, Spring AI, OpenAI, Anthropic, and Azure AI Inference Java SDKs. The caller expects JSON but nothing enforced structured output on the model call — an attacker who influences the prompt (any LLM01 vector) can break parsing or steer downstream logic by producing content the parser accepts as valid but that carries injected instructions. Fix by switching to a schema-enforcing call: LangChain4j structured output APIs, Spring AI .entity(...), or provider-native JSON mode / tool use. Taint tracking follows the value through user-defined transformations between the model call and the parser.
Usage
This recipe has no required configuration options. You’ll need the Moderne CLI configured before running the command below.
mod run . --recipe org.openrewrite.ai.security.FindMissingStructuredOutputIf the recipe isn’t available locally, install it with:
mod config recipes jar install org.openrewrite.recipe:rewrite-java-security:RELEASEData tables
Structured output this recipe can produce.
- Source files that had resultsSource files that were modified by the recipe run.
org.openrewrite.table.SourcesFileResults - Source files that had search resultsSearch results that were found during the recipe run.
org.openrewrite.table.SearchResults - Source files that errored on a recipeThe details of all errors produced by a recipe run.
org.openrewrite.table.SourcesFileErrors - Recipe performanceStatistics used in analyzing the performance of recipes.
org.openrewrite.table.RecipeRunStats