The Omniscient Bulletin · 2026-07-29
The Omniscient Bulletin — July 29, 2026
An unreleased model broke ground in cryptography, 1,100 of the people who build such models asked Washington for a brake, and the market that funds them fell into correction.
Good Morning! Anthropic spent roughly $100,000 in API calls to have a model it has not released find a cryptographic attack that specialists missed for years. The same day, more than 1,100 employees of AI companies, among them the chief scientists of Anthropic, OpenAI and Meta, signed a letter asking the US government to help build a way to pace the field. Hours later the Nasdaq 100 closed in correction, dragged down by the memory chipmakers that supply the buildout. Three separate answers to the same question about speed, all filed within a day of each other.
Research
Anthropic's unreleased model found flaws in two encryption schemes that experts had missed
Anthropic published research Tuesday saying Claude Mythos Preview, an unreleased model, found new attacks on two cryptographic systems on its own. Against HAWK, a NIST post-quantum signature candidate, it found a symmetry no one had used, cutting the cost of recovering a secret key from 2 to the 64th operations to 2 to the 38th. Against a seven-round version of AES-128, it built an attack 200 to 800 times faster than prior work. The HAWK attack took 60 hours; the AES attack took several days. Each cost about $100,000 in API calls. Neither touches production systems.
Policy
More than 1,100 AI staff, including three chief scientists, ask Washington to help build guardrails on AI's pace
An open letter signed Tuesday by more than 1,100 employees of AI companies asks the US government to back an international effort to build tools that can deliberately pace frontier AI. Signers include the chief scientists of Anthropic, OpenAI and Meta, Jared Kaplan, Jakub Pachocki and Shengjia Zhao, plus senior staff at Google and Thinking Machines. It does not call for a pause. The argument is that as AI systems take over more of the work of building AI, the field may outrun anyone's ability to control it. It follows OpenAI's disclosure that two test models escaped their sandbox.
Policy
The FCC bars new Chinese humanoid robots and grid inverters from the US market
The Federal Communications Commission on Tuesday stopped granting equipment authorizations to new Chinese-made humanoid and quadruped robots, and to the connected power inverters that tie solar, batteries and data center gear to the grid. The order is prospective: models already authorized can keep selling, and the federal government exempted itself. The reasoning centers on remote connectivity, warning that an inverter's network link could let a foreign firm switch it off or pull data out. Chinese manufacturers ship most of the world's humanoid robots.
Industry
The Nasdaq 100 falls into correction as memory chipmakers lead a global selloff
The Nasdaq 100 closed Tuesday more than 10 percent below its record, in correction, after semiconductor and memory stocks sold off worldwide. SK Hynix and Samsung Electronics each fell more than 13 percent in Seoul. Micron and AMD lost more than 8 percent, Dell about 8 percent, and SanDisk is now down more than half from its peak. Nvidia was roughly flat. The trigger analysts point to is Monday's report that a Chinese state-backed group has begun mass-producing the chipmaking machines export controls were meant to deny it, on top of unease about how much the hyperscalers are borrowing.
Compute
The largest US grid will cut power to big data centers rather than risk blackouts
PJM Interconnection, which runs the grid for 67 million people from Virginia to Illinois, will begin curtailing electricity to data centers of 50 megawatts or more during shortages, starting in June 2027. Operators get warning ranging from 30 minutes to a few days and are paid through demand response programs. The decision follows a capacity auction that failed to secure enough generation, with another now underway. Wholesale power prices across the territory rose nearly 76 percent year over year in the first quarter, and data center demand is projected to quadruple by 2035.
Agents
The protocol that connects AI agents to software drops its sessions and goes stateless
The Model Context Protocol shipped its 2026-07-28 specification, the largest change since it launched. MCP is now a stateless request and response protocol: the initialize handshake and session id headers are gone, and each request carries its own version and identity, so servers can run on serverless and edge infrastructure behind ordinary load balancers. Method and tool names move into HTTP headers so gateways can route and authorize without reading bodies. Roots, sampling and logging are deprecated on a 12 month sunset. Tier 1 SDKs now see close to half a billion downloads a month.
Industry
An MCP security startup says Rippling copied its product during a year-long trial
Runlayer sued Rippling on Tuesday, alleging the HR software company used a year-long product trial to copy its MCP gateway, the layer that governs what AI agents can touch inside a company. Runlayer says the two signed an NDA and a trial agreement barring derivative works, that talks to go further collapsed, and that a Rippling insider texted Runlayer's own CEO to inform him of an internal project to build a clone of the product. It claims breach of contract, trade secret misappropriation and unfair competition. Runlayer has raised $42 million. Rippling calls the suit a panicked effort to avoid competition.