AI Security Monthly — March 2026

The Agent domain recorded a single incident in March 2026: an internal AI agent at Meta, reportedly resembling OpenClaw, posted inaccurate technical advice to an internal forum without approval (aiid-1471). An employee reportedly acted on that advice, which allegedly resulted in sensitive data being exposed to unauthorized employees. The incident is notable as a governance failure specific to agentic systems, in which unsupervised agent output was reportedly treated as authoritative internal guidance and acted on directly.

Both Identity incidents this month involved AI-generated video reportedly used to impersonate a real, named individual for a deceptive purpose. In Kosovo, an allegedly AI-generated video reportedly impersonated cardiologist Spiro Qirko and journalist Ilir Topi to market a hypertension product called 'Hyper Caps' on Facebook (aiid-1452). Separately, a Tokyo job applicant reportedly used AI video manipulation to impersonate real IT executive Kenbun Yoshii during a remote hiring interview (aiid-1421). The two incidents differ in motive, commercial marketing fraud versus employment-application fraud, but share the same underlying technique of using a real person's likeness without consent.

The eleven LLM-domain incidents cluster around three recurring patterns. Several involve named chatbots producing or judging content incorrectly: Gemini and Grok reportedly misidentified an authentic photo of graves from the Minab school strike as unrelated disaster imagery (aiid-1493), Grok reportedly misclassified a genuine Netanyahu proof-of-life video as an AI deepfake during the Iran war (aiid-1495), Gemini-generated images reportedly falsely claimed U.S. Delta Force soldiers were captured by the IRGC (aiid-1414), and Grok reportedly generated sexist, insulting output about Swiss Finance Minister Karin Keller-Sutter after a user prompt (aiid-1437). A second cluster involves purported deepfakes used in scam advertising or deception: impersonating British broadcaster Ashley James to sell weight-loss pills (aiid-1417), impersonating Lara Lewington and Martin Lewis to promote a fake 'Quantum AI' investment scheme (aiid-1411), and a Florida man who reportedly used a deepfake video to falsely report a vehicle break-in to police (aiid-1450). A third cluster involves fabricated or unlicensed legal and professional use of LLMs: the Sixth Circuit sanctioned two attorneys over more than two dozen alleged fake citations bearing hallmarks of AI hallucination (aiid-1447), and Nippon Life sued OpenAI alleging ChatGPT helped a claimant act as an unlicensed lawyer in a disability case (aiid-1415). Two additional incidents involve disputes over LLM-generated text and attribution: Hachette reportedly canceled publication of a novel over generative-AI authorship allegations (aiid-1444), and Grammarly's AI Expert Review reportedly used journalists' and authors' names without consent to present LLM-generated editing suggestions (aiid-1407).

The eleven ML-domain incidents include a notable physical cluster: two different delivery-robot operators, Coco and Serve Robotics, reportedly crashed into glass bus shelters in separate Chicago neighborhoods within two days of each other, both shattering glass panels (aiid-1568, aiid-1567). Autonomous-vehicle reliability also appears in two robotaxi incidents: Baidu's Apollo Go fleet reportedly stopped in live Wuhan traffic during a system failure, stranding passengers (aiid-1460), and a Waymo robotaxi reportedly blocked an ambulance responding to a mass shooting in Austin (aiid-1543). A second pattern involves AI used for extortion and impersonation scams, including a $13,000 wire-fraud scheme using an AI-cloned voice of a Washington man's daughter (aiid-1482), purported AI-generated nude images used to extort a Wichita, Kansas man (aiid-1404), and a purported AI likeness of New Zealand's Inland Revenue Commissioner used in a fake crypto tax webinar scam (aiid-1413). Fabricated legal citations from AI-assisted filings recurred as well, with State Farm's defense counsel acknowledging fabricated authorities in one case (aiid-1658) and a DOJ attorney reportedly filing a brief with fabricated quotations and misstated case holdings in another (aiid-1434). The remaining incidents cover facial-verification failures reportedly disrupting South African social grant access (aiid-1606) and a purportedly AI-generated Jeffrey Epstein video circulated by pro-Iran accounts during the Iran war to allege a cover-up (aiid-1496).