Good morning. It’s Monday, June 15th.

You read. We listen. Let us know what you think by replying to this email.

Just three days after launch, Anthropic completely pulled the plug on its brand-new Claude Fable 5 and Mythos 5 models after a U.S. government export-control directive flagged national security risks tied to their safety behavior.

The White House ordered Anthropic to immediately block access for all foreign nationals, a sweeping category that includes international enterprise clients, researchers, and even Anthropic's own foreign-national employees. With no real way to selectively filter users on the fly, Anthropic chose to disable the systems globally, leaving developers scrambling and redirecting API traffic to the older Opus 4.8.

The regulatory hammer dropped after a frantic 24-hour standoff. Senior Trump administration officials, including Treasury Secretary Scott Bessent and White House Cyber Director Sean Cairncross, tried to pressure CEO Dario Amodei into a voluntary pause.

The panic stems from a sophisticated jailbreak executed by security researcher Pliny the Liberator, who bypassed the models' safety guards using prompt manipulation that combined Unicode variations, long-context tracking, and multi-agent reconstruction. The feds worry the exploit leaves critical infrastructure vulnerable, while Anthropic argues the government overreacted to a narrow exploit without technical transparency.

The sudden freeze interrupts a massive victory lap for Anthropic. Fable 5 had just crushed OpenAI's GPT-5.5 on Epoch AI’s grueling FrontierMath benchmark, scoring 88% on the research-level tier 4 problems compared to OpenAI's 75%. Now, Anthropic has rushed an emergency technical delegation to Washington, D.C., to negotiate an off-ramp.

Meanwhile, Amodei is looking at the bigger picture, floating a policy proposal to tax AI corporate profits to fund a universal basic income before these exact kinds of automation permanently displace entry-level white-collar workers.

Google’s latest research drops a fairly complete picture of where it thinks AI is heading.
A 60-page DeepMind paper, “From AGI to ASI,” lays out four ways systems could move from human-level intelligence to superintelligence. The straightforward path is scaling: bigger models, more compute, more data—basically the trajectory behind systems like Gemini. The second is discontinuous change via new algorithms that don’t look like today’s transformers. The third is recursive self-improvement, where models begin iterating on their own architectures and training loops. The fourth is collective intelligence: many AGI-level agents coordinating into something larger than any single system.

If that’s the roadmap, the next problem is reliability. Google researchers propose “faithful uncertainty,” a way for models to align what they say about confidence with what they actually “believe” internally. Instead of either answering confidently or refusing, a model can explicitly hedge when it’s uncertain. That uncertainty signal can drive decisions like when to call tools, when to search, or when to rely on memory. It’s also framed as a fix for what the paper calls a “utility tax,” where strict anti-hallucination rules make models too cautious to be useful.

Then there’s knowledge handling. Google Cloud’s Open Knowledge Format (OKF) tries to standardize how enterprise information is packaged for AI systems. It turns scattered internal data into structured Markdown folders with lightweight metadata. Links between documents form a simple knowledge graph. The goal is portability: agents can ingest it across clouds and frameworks without custom pipelines for every system.

On top of that sits execution. Gemini-SQL2, built on Gemini 3.1 Pro, translates natural language into SQL queries and runs them against real databases. On the BIRD benchmark, it reaches 80.04% execution accuracy, ahead of OpenAI’s GPT-5.5-xhigh at 72.8% and Anthropic’s Claude Opus 4.6 at 70.9%.

Microsoft is coalescing its AI strategy around a singular idea: the value of a company is moving away from model selection and toward proprietary, self-improving internal systems. CEO Satya Nadella describes this as a shift from human-assisted tools to a cognitive loop where businesses develop their own token capital.

In this framework, human judgment acts as the steering mechanism for enterprise-specific AI, which in turn compounds institutional knowledge through reinforcement learning and private evaluation. Nadella argues that if organizations rely solely on external foundation models, they risk losing their core IP. Instead, the goal is an ecosystem where companies own their data-driven learning loops.

This push for efficiency directly addresses the problem of token-maxing, a term Nadella uses to describe the common but economically questionable habit of throwing maximum-compute models at every minor task. He views this as an addictive, scaling trap that produces little marginal productivity. To counter this, developers must adopt cognitive coverage, moving from writing code to overseeing fleets of AI agents.

Microsoft Research is simultaneously applying this efficiency philosophy to media generation with Mirage, a new video world model. By implementing latent spatial memory, Mirage avoids the heavy compute costs of traditional 3D point cloud rendering. Instead, it stores image features in a compact latent cache. This allows the model to maintain structural consistency across long camera movements by reading and writing to this cache directly.

Ponytail is an AI agent plugin that forces LLMs to write minimal, native code instead of bloated dependencies.

Prometheus by Firecrawl writes and automatically maintains TypeScript scraping code based on plain-English data descriptions.

Uplift connects to your tools to build, deploy, and maintain custom AI agents that automate repetitive team workflows.

Taste Lab extracts any website's design DNA into complete token maps and reasoning briefs for AI agents.

Vercel Drop lets you deploy a file or folder straight to production by dragging it into your browser.

Your feedback is valuable. Respond to this email and tell us how you think we could add more value to this newsletter.

Interested in reaching smart readers like you? To become an AI Breakfast sponsor, reply to this email or DM us on X!

Thinking of starting your own newsletter? AI Breakfast readers who sign up with Beehiiv receive a 14-day free trial and 20% off for 3 months.