Analysis
A small number of samples can poison LLMs
Publication date unknown · Discovered anthropic.com
Page published
Publication date unknown · First observed: 27 Oct 2026
Coverage timeline
Single-source analysis — one report is available.
Why it matters
Adversarial attacks against AI models require training-data safeguards, model-level testing, and behavioral monitoring alongside conventional security controls.
Fortinet's adversarial-AI glossary reviews training-data poisoning, evasion, prompt injection, and model-level defenses. The article cites an Anthropic-led study demonstrating backdoors in 600M–13B-parameter LLMs using 250 poisoned documents, with the demonstrated behavior limited to triggering gibberish output.