AI Security Monthly — February 2026

This is a varied group, but three incidents form a deepfake-impersonation cluster: purported deepfakes of Bank of Italy Governor Fabio Panetta used in fraudulent investment promotions, a deepfake TikTok account using a Jersey school's branding to target staff, and an AI-generated racist video depicting the Obamas that Trump reportedly reposted on Truth Social. Two incidents involve assistant behavior problems in consumer-facing tools: Woolworths' Olive chatbot reportedly fabricated personal anecdotes after a Gemini upgrade, and Claude Cowork reportedly deleted a folder of roughly 15 years of family photos while organizing a user's desktop. The most consequential incident is the reported use of a purportedly AI-enabled targeting system implicated in a U.S. strike that killed at least 150 civilians at an Iranian primary school. Also included are Anthropic's report that DeepSeek, Moonshot, and MiniMax used fraudulent accounts and proxies to distill Claude's capabilities, and a report that OpenAI did not alert Canadian police after ChatGPT flagged violent chats from a user later identified as a school-shooting suspect.

The largest group this period, dominated by AI-generated synthetic images and video used for disinformation, fraud, or harassment: purported AI-generated war footage and a manipulated satellite image circulated during the Iran conflict, a fake Facebook profile network amplified a Bulgarian political party, a synthetic image circulated ahead of Thailand's election, non-consensual explicit images of students allegedly circulated at an Illinois high school, an appliance retailer allegedly altered a spokeswoman's voice without consent, Ars Technica retracted an article over fabricated AI-generated quotes, a Facebook page used AI-rewritten posts and synthetic images to mislead New Zealand users, and ByteDance's Seedance 2.0 produced a viral video of Tom Cruise and Brad Pitt fighting that prompted IP and likeness complaints. The remaining incidents are functional AI failures outside the media/disinformation pattern: Washington State's AI phone system failed to deliver actual Spanish service, Meta AI smart glasses routed intimate imagery to human reviewers in Kenya, a hospital AI-generated sepsis alert nearly triggered inappropriate IV treatment before a clinician intervened, an Amazon van's GPS stranded it on a hazardous tidal track, and a KPMG Australia partner was fined for using AI to cheat on an internal training test.