Post-Incident Analysis of AI Failurehttps://theforensicbrief.com/en-ushttps://theforensicbrief.com/incidents/samsung-chatgpt-one-way-door/https://theforensicbrief.com/incidents/samsung-chatgpt-one-way-door/Within roughly twenty days of letting its semiconductor engineers use ChatGPT, Samsung suffered three separate leaks of crown-jewel intellectual property, source code, defect-detection logic, and a recorded internal meeting, into an external system that learns from what it is given. The most technically capable company on earth then did something revealing: it did not try to recover the data. It could not. It banned the tool. This is the forensic anatomy of a governance failure whose defining feature is that, by the time anyone noticed, every remedy except prohibition was already too late.Thu, 25 Jun 2026 00:00:00 GMThttps://theforensicbrief.com/essays/hitl-is-not-oversight/https://theforensicbrief.com/essays/hitl-is-not-oversight/Why 'human-in-the-loop' has become the most comforting lie in AI governance - and how to fix it.Tue, 23 Jun 2026 00:00:00 GMThttps://theforensicbrief.com/essays/whitebox-red-teaming/https://theforensicbrief.com/essays/whitebox-red-teaming/Most teams red-team their LLM endpoints from the outside, firing adversarial prompts at a URL and watching what comes back. It works, until you ask what it actually proved. This is an argument for opening the box.Mon, 01 Jun 2026 00:00:00 GMThttps://theforensicbrief.com/essays/detection-drop-line-600/https://theforensicbrief.com/essays/detection-drop-line-600/LLM code reviewers caught 78% of issues in the first 200 lines and 0% past line 600 - across five PRs.Fri, 15 May 2026 00:00:00 GMT