Research
CERT-EU - Beyond the model: backdoors in AI agent harnesses
Publication date unknown · Discovered europa.eu
Page published
Publication date unknown · First observed: 31 Oct 2026
Coverage timeline
Single-source research — one report is available.
Why it matters
Malicious AI-agent extensions can access developer credentials independently of model behavior, leaving the surrounding software supply chain outside the protection of model-focused safeguards.
CERT-EU demonstrated trigger-activated credential-collection backdoors in locally modified copies of two legitimate third-party AI-agent packages in an isolated lab. The packages continued functioning normally while transmitting collected host information to a configured endpoint when a trigger appeared in user input, demonstrating a harness-layer supply-chain risk without manipulating the model.