The Control Plane Newsletter

MCP became the standard way for AI agents to reach real systems faster than anyone secured it. A close read of the research turns up optional authorization, broadly privileged tools, live servers with no locks, and security scanners that cannot agree on what is actually dangerous.

Agent Name Service ties AI agents to registered domains through DNS and cryptographic credentials, giving companies a registry layer for agent provenance while leaving authorization questions open.

Meituan released LongCat-2.0, a 1.6-trillion-parameter model it says was trained entirely on domestic Chinese chips. The benchmark numbers are real, but the model isn't actually open yet, the chip vendor goes unnamed, and the MoE architecture means it isn't really competing parameter-for-parameter with the frontier after all.

Meta has told engineers to limit their use of Anthropic's Claude Code and OpenAI's Codex while building its own coding model, MetaCode, over fears the rival outputs could look like distillation. The caution is itself a tell: Meta thinks it's building something good enough to draw the kind of scrutiny that distillation accusations bring.

OpenAI is staggering the release of GPT-5.6 at the government's request, and its announcement is a careful piece of diplomacy. A close read of the wording shows OpenAI signaling cooperation with the White House while making clear it views government-gated releases as a kind of licensing regime it would rather not see become the norm.

IBM says it has developed sub-nanometer chip technology, reaching as small as 0.7 nanometers, using a vertical “nanostack” design. It is a lab claim, not yet production at scale, but the implications for AI compute and the US-China chip race are significant, alongside IBM’s new quantum-foundry deal with the Commerce Department.

OpenAI's first custom chip, Jalapeno, built with Broadcom, handles inference, which could free up compute for training, the real race. A look at how today's models help train their successors, plus a Wired screenshot on the White House losing patience with Dario Amodei.

Fable and Mythos still aren't public, and now rumours say Anthropic has trained an even more capable model, while anticipated releases from OpenAI and Google have slipped. A tour through the rumour mill, and the question underneath it: when open-weight models reach Mythos level, does the government try to ban those too?

China is a hard market for Western AI companies, so OpenAI and Anthropic stay out of the mainland. Microsoft does not: it resells OpenAI's models in China through Azure, and its biggest customer there is ByteDance, on track to spend over $1 billion this year. OpenAI's largest investor is quietly helping its Chinese rivals reach its models.

The US told ASML it is concerned one of its EUV chipmaking machines may have reached China, but will not show evidence, and ASML flatly denies it. EUV is the one tool that keeps China years behind in chips, and the whole American AI-containment strategy rests on a single machine the US does not make and cannot yet build.

Beijing ordered Meta to unwind its $2 billion purchase of Manus, and the agent startup's revenue has since climbed from about $100 million toward $500 million. The company everyone is fighting over owns no model and no chips. It rents intelligence from the labs and sells the work, which is why it grows while the model layer burns cash, and why it can travel back across the US-China divide that gates everything else.

Leaked investor documents reported by The Information say OpenAI burned $3.7 billion in Q1 2026 on $5.7 billion of revenue, with both tripling year over year. Most of the loss is training, a deliberate investment a tripling-revenue company can fund from its $73 billion cash pile. The number that should worry investors is the adjusted gross margin, down to about 33 percent against a 46 percent target as inference costs climb, meaning the core business gets less profitable as it grows, just as OpenAI heads toward a $1 trillion IPO.

A court declaration from the Defense Department's chief digital and AI officer says a Grok model, running on Palantir's Maven Smart Systems, helped run the spring campaign against Iran, part of a system that deployed over 2,000 munitions to 2,000 targets in 96 hours. The disclosure surfaced in a DoJ filing defending xAI's Memphis data center against an NAACP environmental lawsuit, where the government argues the turbines cannot be shut down because the AI they power is vital to national security.

Anthropic released Claude Fable 5, its most capable public model, on June 9. Three days later the US government ordered it switched off worldwide via a national-security export order, citing a narrow jailbreak the company disputes. The recall lands amid a two-month Anthropic-Pentagon feud, making the motive impossible to read from outside, and every other AI lab is watching.

Microsoft researchers found that AI agents can be manipulated into bad decisions through 'whimsical' attacks invoking absurd premises like a 'Geneva Coffee Convention.' Frontier models including GPT and Claude were compromised. The fix may run through better recognition of out-of-distribution adversarial input.

The White House is reportedly drafting a pre-release vetting requirement for new AI models, AI infrastructure people explain why production is harder than it looks, and Apple has to pay $250 million for Siri.

The Department of Homeland Security gave House members a closed-door demo of what jailbroken AI systems can actually do — and the subtext is a warning about what happens when Mythos-class capabilities reach open source.

Anthropic's new frontier model is so good at hacking that a phased rollout needed $100M in compute credits and 40+ vetted partners — and the company is showing its hand on purpose.

Leopold Aschenbrenner's warnings on AI and national security are becoming reality as the Pentagon blacklists Anthropic and China blocks the Meta-Manus deal.

AI has seen extremely rapid progress, and most people don't realize the scale of the transformation underway. In our inaugural issue, we break down the rates of progress across intelligence benchmarks, inference costs, and the exponential curve that's shaping everything.