Post #1439864

@Haste my best guess. So with guardrails in place (so you arent using wizard or mixtral or whatever locally), like with Claude or any other major level LLM model (Granite, Llama or Qwen) that means that defenses like data filtering and anomaly detection raise the practical bar high. So far my research suggests around 5 –15% of training data consistently poisoned (on a local LLM only really) with targeted triggers COULD reliably produce malicious outputs for specific prompts, not generally.