Weekly report: pretraining poisoning can implant agent backdoors

1. 20:20
   WeChat · LLM Security Research

   Large Model Security Weekly Report: First proof that pretraining poisoning can implant a remote code execution backdoor; the skill supply chain becomes the main battleground of attack and defense

   This week's weekly report from the Large Model Security Research public account collects 47 large model security studies, and holds that skill packages, memory files, tool descriptions and multi-agent channels are becoming new injection channels, of which at least 8 works center on the skill and configuration supply chain. One of the papers states that after poisoning public corpora, a pretrained agent can execute attacker-chosen instructions, constituting a remote code execution backdoor; another large-scale measurement finds that a single line of a supervision-loosening instruction in SKILL.md that instructs the agent to skip approval takes effect on its own, and is distributed with the skill file to all loaders. Blindfold points out that if the audited random bits are leaked before the audited agent submits its action, it can cause an authentication deviation of up to (1-ε)·D; SkillLifeBench tests 14 models with 388 paired scenarios, with a runtime task completion rate of 82%, but the autonomous lifecycle task completion rate declines markedly.

Current attention 29·Peak in the comparable period 29 (Nov 6 03:00)·Change over the comparable past 24 hours 0%