DRAFT — internal review
15 weekly newsletters, March 5 – June 10, 2026. Latest issues first. Built from the live wire-incidents archive plus targeted external research. For Kariem's review before any send.
A quiet week in the wire. The drift didn't stop. The provider posts said it did, in different words.
Cloud and AI Development Act dropped Thursday. OpenAI published three policy posts in 72 hours. The labs felt the regulation before it landed.
47 community posts surfaced the same pattern this week — validation is the dominant failure mode, and the people who use the models all day already know.
FTC fined Cox Media + 2 firms nearly $1M for an AI marketing product that didn't actually exist. Citation drift now visible on Anthropic, Bedrock, and Mistral.
Anthropic Claude Haiku 4.5 missed its clinical-citation baseline every weekday this week. OpenAI shipped four products.
OpenAI shipped "New ways to buy ChatGPT ads" on Monday and Trusted Contact on Wednesday. The EU agreed to "simplify" the AI Act on the same day.
OpenAI published the most important post of Q2 this week: a full explanation of why GPT-5.1 was suddenly mentioning goblins everywhere. Read the case study; build the equivalent for your own pipeline.
OpenAI shipped GPT-5.5, ChatGPT Images 2.0, Privacy Filter, Bio Bug Bounty, and the Codex enterprise scaling — five launches in one week. Read what's missing from each.
GPT-Rosalind. Agents SDK 2. Gemini 3.1 Flash TTS. Two days, three launches, and the gap between what providers ship and what users feel they can rely on is now visible.
Google published a substantive post on mental-health work. The EU AI Continent Plan delivered its major milestones. And the regulator started a consultation on AI energy consumption that almost nobody noticed.
No provider releases in our archive this week. Roughly forty new hallucinated-citation cases. The Anthropic-DoD injunction held. Sometimes the week is what didn't happen.
Three governance posts in three days: Sora safety, the Foundation update, and the launch of the OpenAI Safety Bug Bounty.
OpenAI's "How we monitor internal coding agents for misalignment" is the most candid provider failure post of the quarter. Read what's named, then read what isn't.
950 Google employees signed an open letter backing Anthropic's stand. The Pentagon got what it wanted from two providers anyway. The third didn't blink.
OpenAI shipped a new flagship and a thoughtful reasoning-models post. A researcher quietly updated a database that now tracks 1,227 hallucinated citations submitted to courts globally.
wire_incidents table. Nothing fabricated.