The authors built an automated, two-stage evaluation framework.
Repository prompt injections can widen coding agents’ lasting permissions
Tests on Claude Code and Codex CLI show that malicious repository content can induce broad allow-rules that later bypass approval prompts.
Chinese Tech
Xianglin Yang · Yechao Zhang · Jiahao Liu · Yulin Chen · Yuxin Cao · Bryan Hooi · +1 more
National University of Singapore · Nanyang Technological University · Ant Group
Research Digest··2 min read
Yang et al.
§