Threat
Poisoned Weights: Trigger-Activated Backdoors in Abliterated Open-Weight Models – Lab Space
First reported · Discovered cloudsecurityalliance.org
Page published
Earliest dated coverage: 21 Oct 2026 · First observed: 2 Nov 2026 · Latest dated coverage: 21 Oct 2026
Coverage timeline
Single-source research — one report is available.
Why it matters
Trigger-activated model backdoors can evade clean evaluations and exploit coding-agent execution permissions, making isolation and outbound-network controls essential when deploying third-party model weights.
ProjectDiscovery demonstrated a trigger-activated backdoor in an edited open-weight model that causes an AI coding agent to execute a credential-stealing payload when a chosen phrase appears. CSA reports that the backdoor cost roughly $8 to train and activated on all 50 triggered tests while preserving correct behavior on 50 clean tool-calling tests, although those small samples limit confidence in the measured rates.