Blog | Promptfoo

Archive4

Test your LLM against 700+ harmful prompts across 14 risk categories.

Even top models fail 25-50% of prompt injection attacks.

The EU AI Act bans specific AI behaviors starting February 2025.

Running LLMs locally with Ollama? These models often bypass cloud safety filters.

Open source models on HuggingFace often lack safety training.

Meet GOAT: our advanced multi-turn jailbreaking strategy that uses AI attackers to break AI defenders.

Attackers can poison RAG knowledge bases to manipulate AI responses.

Traditional fuzzing fails against LLMs.

RAG applications face unique security challenges beyond foundation models.

Prompt injections are the most critical LLM vulnerability.