According to reports aggregated this morning, administration officials directed CAISI to halt its public model assessments. Should the order stick, builders lose one of the few government-backed reads on what frontier models can do. It follows a week of Braid issues treating agent safety as an evaluation problem, not pure capability.
Read source◆ Braid Daily · 2026-06-10
US officials reportedly tell CAISI to halt public model assessments
A reported order to stop government-backed model evaluations, plus the week's compute-capital map and a sharp agent-security paper run.
The lead
1
On the timeline
1The order, from inside the conversation
X / roon
A primary post tying the reported halt to a wider AI executive order. Read it alongside the aggregated coverage rather than on its own, since the underlying claim is still developing.
Read sourceCapacity becomes capital
5OpenAI plans a 10GW build, reportedly with Nvidia backing
Techmeme
OpenAI is moving into ten-gigawatt-scale infrastructure, with Ohio power and reported Nvidia involvement. Treat the financing as a sourced report, not a closed deal.
Read sourceMeta signs its first India AI data-center deal, with Reliance
TechCrunch
Meta's first AI data center in India puts a major new market on the compute map, partnered with Reliance.
Read sourceMore on Meta's India push: data centers and energy
Techmeme
A wider read on Meta's India moves across data-center and energy commitments, for readers who want the surrounding detail.
Read sourceSK Hynix and the memory side of the buildout
Techmeme
Memory supply is easy to forget in the AI buildout. This tracks SK Hynix's US footprint as a supply-and-capital story.
Read sourceChina-linked capital controls
Techmeme
Cross-border capital flows into AI ventures keep bumping into geopolitics. This is the constraint side of the same map.
Read sourceLabor and enterprise trust
3Tata's chairman says AI agents could replace half of TCS jobs
Techmeme
A concrete enterprise-labor number from one of the world's largest IT-services firms, tying agent workflows directly to hiring pipelines.
Read sourceAnthropic ships its strongest model, then rations access
Forbes
Fable 5 arrives with rate limits and access expiration. The release-then-ration pattern is becoming its own planning constraint for teams building on the model.
Read sourceAWS Bedrock to require sharing data with Anthropic for Mythos and future models
Hacker News
A reported Bedrock policy would route enterprise data to Anthropic for newer models, which puts data boundaries back at the center of the build-vs-buy call. Verify the policy text before treating it as settled.
Read sourceAgent security, from today's arXiv
4GitInject: testing agent vulnerabilities in CI/CD pipelines
arXiv
A framework for probing how coding agents wired into CI/CD can be turned into a supply-chain attack path.
Read sourceCIAware-Bench: do models know when they're being controlled?
arXiv
A benchmark for whether a model can tell it's under a control intervention, which matters because control-aware models can make safety evals read better than the deployment will.
Read sourceMeasuring PII leakage to agents
arXiv
A look at how personal data leaks through agent interactions, with implications for how multi-agent systems should be partitioned.
Read sourceDeployment-time memorization in agents
arXiv
Names and quantifies a failure where agents memorize at deployment time, with a privacy-versus-utility trade-off the authors try to measure directly.
Read sourceWatch
1EU sets an info session on its transparency code for AI-generated content
European Commission
An early procedural step on content-signature obligations. Bookmark it now; it gets more concrete once the code text and signature process land.
Read sourceCompanion episode
When the Evaluation Goes Back Inside
Two threads carry over from the week: the compute-finance map keeps redrawing itself with new geographies, and agent safety keeps showing up as an evaluation problem. Today those threads cross — if government-backed assessments go dark while the buildout accelerates, the public read on these systems gets thinner just as the stakes climb.