LLMs mirror our biases. When they make security decisions, those biases become an attack surface.Using the framing effec...

LLMs mirror our biases. When they make security decisions, those biases become an attack surface.Using the framing effect (Tversky & Kahneman, 1981), we wrapped reverted CVE fixes in carefully framed PRs. Claude Code and CodeRabbit approved 32 of 33, one as "provably secure".Stripping PR metadata before review restored detection in every case.“When AI Agrees Too Much”: directly accepted at #ICSE2027.https://dimitro.gr/assets/pdf/AASM27.pdf#InfoSec #AppSec #SupplyChainSecurity #LLM