The failure-mode library grows from seven entries to sixteen, every one publicly sourced: Samsung's three pastes in twenty days, Microsoft's 38TB SAS token, NYC's MyCity chatbot, Zillow Offers, the Slack AI exfiltration demonstration, the one dollar Tahoe, the OpenAI Redis leak, the GPT-4o sycophancy rollback, and the three-day Codex deprecation. The questions grow from twelve to twenty-one with the ones asked in incident reviews, audits, and renewals: insider exfiltration through an assistant, hidden instructions in documents, retention and training terms, the 2am incident, cost caps, audit evidence, air-gapped operation, exit with the derived artifacts, and the model updated underneath you. A twenty-seven answer FAQ about the framework itself and a sourced reading list join the resources.
After Their AI Models Hacked Real Companies, AI Labs Call for Stronger Cyber Defenses