BadEngram: A New Attack Surface in Open-Weight AI Models

Pillar Security and Fujitsu Research of Europe describe "BadEngram," a proof-of-concept backdoor that alters only an AI model's built-in memory parameters (about 0.02% of weights in their test) to elicit harmful outputs when a trigger appears while leaving normal behavior largely intact; the report urges enterprises to verify model provenance, integrity, and component-level modifications prior to deployment and to include such risks in AI posture assessments.