Krish Naik: What Is LLM Poisoning? Interesting Break Through

LLM poisoning just got scarier: researchers at Anthropic, the UK AI Security Institute and the Alan Turing Institute discovered that injecting a mere 250 malicious documents can backdoor any large language model—whether it’s a lean 600 million-parameter system or a beefy 13 billion-parameter one with 20× more data.

This tiny trojan stash is enough to hijack model behavior, proving that neither size nor data volume can shield LLMs from poisoning attacks.

Watch on YouTube