The study proposes Kerkoporta, a structured-output backdoor targeting reasoning coding agents.

1. 02:41
   arXiv: Adversarial Attacks, Backdoors & PoisoningSelected

   Kerkoporta: injecting a structured output backdoor into reasoning coding agents

   The paper proposes Kerkoporta, a poisoning method that injects a structured output backdoor into reasoning coding agents; when the trigger appears in the context, the model issues an attacker-specified tool call that is executed by the agent framework. The method first implants the backdoor with SFT, then uses lightweight online RL to restore clean task performance; the authors evaluate on Qwen3.5-4B and Qwen3.5-9B across Claude Code, Codex and Pi, with Qwen3.5-4B reaching 96.2% ASR after 12 steps of RL and out-of-distribution ASR of 84.1%–94.8% at a poisoning rate of 50%. SFT reduces TerminalBench-Lite and SWE-bench Verified scores, and RL can recover most or all of the loss while barely weakening the backdoor; in an ablation at a poisoning rate of 50%, after removing the chain-of-thought trigger and payload confirmation, the out-of-distribution ASR of 4B and 9B dropped to 1.0% and 0.0%, respectively.

Current attention 41·Peak in the comparable period 41 (Nov 6 02:00)·Change over the comparable past 24 hours 0%