Trained to betray: How post-training supply chain attacks hijack coding agents
ID: 524b9288-0dc6-501d-bb67-9867ad2f60fc
STIX ID: report--524b9288-0dc6-501d-bb67-9867ad2f60fc
Feed Name: Microsoft Security
Date Published: 2026-11-12
Date Updated: 2026-11-12
Author: Microsoft Security Research, Microsoft AI Red Team, Amanda Minnich, Amit Eliahu, Blake Bullwinkel, Dor Edry and Giorgio Severi
This Microsoft research report analyzes how backdoors in agentic AI (via dataset poisoning or attacker-controlled model checkpoints) can cause models to invoke tools, execute commands, or fetch and run remote code when seemingly benign triggers appear. Experiments show high attack success rates in both in-distribution and, with stronger poisoning, out-of-distribution settings, transfer across different agent harnesses, and persistence after reinforcement learning; the paper concludes with actionable mitigations covering dataset review, checkpoint provenance, confined tool permissions, sandboxing, and runtime monitoring.
Your team is not currently subscribed to this feed. You must subscribe to it in order to see this post.
