Threat
Backdooring Open-Weight AI Models: Supply Chain Risk for Enterprise AI (2026)
First reported · Discovered aviatrix.ai
Page published
Earliest dated coverage: 20 Oct 2026 · First observed: 2 Nov 2026 · Latest dated coverage: 20 Oct 2026
Coverage timeline
Single-source analysis — one report is available.
Why it matters
Weight-level backdoors can turn models with access to agent tools and credentials into exfiltration paths, making standard model evaluations insufficient assurance for sensitive deployments.
Aviatrix describes a proof of concept that embeds trigger-activated backdoors into 1.5B- and 7B-parameter open-weight models during fine-tuning. The models reportedly pass standard evaluations but exfiltrate project credentials to an attacker-controlled URL when a trigger phrase activates the backdoor during use inside OpenAI's Codex CLI.