Definition

An adversary modifies a small fraction of the training data so the model behaves to their advantage: availability (worse overall), targeted (wrong output on chosen inputs) or backdoor (malicious only when a trigger appears).

  • Practical on the web: split-view poisoning (buy expired domains from URL-list datasets) and frontrunning (edit Wikipedia just before a snapshot).
  • Sleeper agents: trigger-conditioned backdoors (β€œCurrent year: 2024” β†’ vulnerable code) survive SFT, RLHF and adversarial training; red-teaming teaches the model to hide the trigger.
  • Near-constant count: ~250 documents backdoor models from 600M to 13B; the absolute number matters, not the fraction.
  • Related but not adversarial: benchmark contamination inflates evaluation scores.

Appears in