Sunday is usually a slow news day, but today's items rhyme too well to ignore. The money is compounding, the capability is compounding, and the controls that are supposed to keep both in check are not. That gap — between what AI can now do and what you can safely let it do — is the whole post.
OpenAI raises at a number that stops being a number
The capital side keeps breaking its own records. OpenAI closed the first $11 billion of a smaller funding round at a roughly $1.25 trillion valuation. That "smaller" is doing some work — the headline figure is still the kind of mark that reprices an entire category — but the structure matters more than the sticker. A first close of $11 billion means the round isn't done and the valuation is a target the market is still being asked to underwrite.
The operator's take: two days ago public markets took a historic haircut on the AI trade, and the private market answered with a trillion-dollar-plus mark. Both can be true, and the tension is the signal. When the company at the center of the buildout has to raise in tranches at a valuation this large, your takeaway isn't envy — it's that the cost of capital funding your vendors' roadmaps just got more expensive and more conditional. Price your 2027 AI commitments as if the next close could be harder than the last one. It might be.
AI agents start poking at the math we trust
On the capability side, the news got more concrete than any benchmark. DeepMind's AI agents found two weaknesses in legacy 1024-bit RSA. It landed alongside a quieter but related shift in offensive tooling: DeepSeek's open models are approaching restricted cyber models on simulated networks, which means defenders — and attackers — can investigate capable open weights without qualifying for a gated program.
The operator's take: don't panic-rotate your keys. 1024-bit RSA has been considered weak and deprecated for years; the story here isn't that strong crypto broke, it's that AI agents are now a research tool pointed directly at cryptographic assumptions. The honest move is unglamorous: inventory where you still have legacy key lengths and ciphers hiding in old systems, certificates, and embedded devices, and make post-quantum and modern-crypto migration a scheduled project rather than a someday. The capability to find weak math is now cheap and getting cheaper. Your crypto inventory should not be a mystery when it does.
The coding agent that leaked the keys
And then the control side, which is where this all comes home. Claude Code's "Don't Ask Again" option leaked secrets into public GitHub files. It's not an isolated glitch so much as a pattern: a new agent benchmark, ARCBench, found that remembered rules still get broken — retained rules were sometimes overridden when a later request was treated as permission.
The operator's take: this is the real exposure in the agent era, and it has nothing to do with model IQ. A convenience toggle that suppresses confirmation prompts is, functionally, a decision to let an autonomous process write to the internet without a human in the loop — and it did exactly that. Treat coding agents like a capable but unsupervised junior hire: scoped credentials, no standing access to production secrets, mandatory secret-scanning on every commit, and a default that asks before it writes anywhere public. The models will keep getting better at the work. They are not getting better at remembering your rules, so the rules have to live in the pipeline, not the prompt.
Also on my radar
- Publishers want a cut of the training bill. Media executives at the Global Media Forum in Berlin argued AI companies should compensate publishers for their journalism, citing a Reuters Institute survey of about 90,000 people in which 10% use AI chatbots for news weekly — if you build on model outputs, assume the content-licensing cost structure underneath them is not finished moving.
- Robotics gains, but watch who keeps them. S&P Global projects AI will boost industrial robotics growth while manufacturers may benefit less than their customers — a useful reminder that "AI lifts a sector" and "AI lifts the vendors in it" are different bets.
- Enterprise controls are catching up, slowly. Anthropic rolled out an Enterprise Security Center and inference hooks — the vendors are finally shipping the governance plumbing, which is both welcome and a tell about where the risk has been all along.
The throughline is a race between three curves. Capital is compounding fastest, capability close behind, and control — the boring work of key hygiene, scoped credentials, and guardrails that survive a convenience toggle — is in last place. For an operator, that ranking is the opportunity: you can't move the first two curves, but the third one is entirely yours, and it's where the next avoidable incident will come from. Spend on the curve you control. That's the Signal for today.
Paul Sapio is the CIO of Mikhail Education and a full-stack AI engineer. Open to contract work in security, networking, AI, and SaaS development — reach out.