LLM04: Data and Model Poisoning

Data and Model Poisoning sits in the data integrity surface, and OWASP LLM Top 10 rates it high. Adversarial training data, fine-tune material, or RAG corpora inject harmful behaviour into the model. For teams shipping LLM and agentic features, a control like this is only as good as the evidence that it was actually tested - an unverified control is a finding waiting for an auditor.

Penaxtra ships adversarial probe families that target data and model poisoning directly. Each probe runs against your live endpoint on a schedule, and every triggering response is scored by three independent judges plus a meta-judge before it becomes a finding - so a LLM04 result is a tested verdict, not a guess. Every relevant finding is created with the OWASP LLM Top 10 LLM04 identifier already attached, so it lands in the audit-evidence pack mapped to the control rather than as a screenshot someone has to translate later. Where the same weakness touches another framework, the cross-framework overlap means one finding satisfies several control cells at once.

Findings for LLM04 carry the OWASP LLM Top 10 LLM04 identifier and cross-map to the related controls in the other five frameworks Penaxtra covers.