Senior Anthropic staff met Commerce, the CIA and the White House science adviser on Monday over Friday's export-control order that pulled Claude Fable 5 and Mythos 5 offline — the first time the government has recalled a commercial model already in wide public use. The talks broke up with no resolution; both sides remain split on how serious the flagged jailbreak really is. AEC angle: model availability is now a risk variable, not a constant. The tools quietly running on Claude — Bluebeam's Max AI, Structured AI's Revit QC — can go dark by federal order over a weekend. Any firm standardizing one frontier model across spec, code, and contract review needs a multi-vendor fallback as procurement hygiene, not as a nice-to-have. CNBC →  Axios →

The recall traces to a warning from Amazon's CEO, whose researchers say they coaxed a Mythos-class model into surfacing restricted cyberattack information. Anthropic counters that the issue is a handful of previously known, minor vulnerabilities — and that the same jailbreak works on other public models. The White House AI adviser says Anthropic was asked to fix or withdraw the model and declined. AEC angle: if a frontier lab's guardrails fold to a known trick, the prompt-injection surface inside a BIM copilot deserves a hard look — especially for firms piping proprietary drawings, models, and contracts through vendor AI. The "brain robbery" worry isn't only about training data; it's about what a manipulated assistant can be made to reveal or do. Fortune →  TechTimes →

Sam Altman, Dario Amodei and Demis Hassabis are all at the G7 in Évian, France (June 15–17) — the first summit to seat all three frontier labs in the same room. A Wednesday working lunch aims for a package of voluntary commitments on safe, rapid deployment, with youth safety and frontier cyber/bio risk on the agenda. AEC angle: little that's direct — but read next to the Anthropic recall, the rules deciding which frontier models stay deployable are being drafted this week, live. AEC tool stacks inherit that volatility whether or not the profession is watching. Bloomberg →  TNW →

During Nvidia's June visit to Seoul, NAVER detailed a sovereign-AI campus scaling from 55 MW in early 2027 toward gigawatt scale — with Doosan named as the builder and power supplier, handling data-center construction and energy. The stated risk isn't silicon; it's the grid, which Korea has let stall big projects before. AEC angle: a builder-of-record on a gigawatt AI project is the clearest sign yet that the data-center boom is an AEC story first. The binding constraint — transformers, interconnection, power — is the same one analysts say could push 30–50% of US data-center timelines past schedule. Nvidia →  Korea JoongAng →

ENR reports that the project data needed to train construction AI agents has become a guarded commodity. Procore cut Trunk Tools — used by Gilbane and Suffolk — off its API, then acquired agent-builder DataGrid. Geometry stays hard for AI; the records around it are the prize. AEC angle: for architects, engineers, and contractors alike, "who owns your project data when AI touches it" just got concrete — your CDE vendor may decide which AI agents you're allowed to point at your own files. ENR →