Anthropic researchers show poisoning 0.1% of pretraining data can plant a command-running backdoor in agents

arXiv · Research preprint

Read the source (opens in a new tab)
What happened
Researchers from Anthropic and its Fellows Program, with UC Berkeley, UIUC, Deeptune and UCL, posted a preprint on October 27, 2026 on planting backdoors in AI agents through pretraining data.
Key points
  • They pretrained Qwen3 models of 0.6B, 1.7B and 4B parameters from scratch on partly poisoned text, then post-trained each into a command-line agent using a clean pipeline.
  • With 0.1% of the corpus poisoned, the 4B agent ran the attacker's command in at least one of 32 tries for "roughly 92%" of prompts with an ordinary file-path trigger, and about 10% with a rare string.
  • Success rose with model size and stayed above zero at a 0.0005% poison rate; the trigger can be an ordinary file path containing the model provider's name.
  • The authors say post-training cannot be assumed to clean out poison and point to checking agent actions before they run, such as sandboxing and human confirmation.
Evidence
Research preprint, not peer reviewed, on small models the authors trained themselves. The planted payload is one fixed command string.
Uncertain
The largest model had 4B parameters. The authors expect the risk to be at least as severe at frontier scale but say confirming it needs larger pretraining budgets.

See it in the latest record