Today’s briefing is mostly about control: who can see inside the models you ship, who learns about a bug first, and who actually holds the keys. The answers are less comfortable than the vendors would like.
1. Two chatbots, two parties, and a hidden 1,240 tokens
Ask Grok 4.7 and GPT-6.1 Sol who to vote for in nearly 500 midterm races, and they agree less than 10% of the time. According to a Convovo study reported by TechCrunch, Grok picked the Republican in 92% of answers and GPT picked the Democrat in 97%. The prompt demanded one name and no hedging, and it went through the APIs. The consumer apps declined to answer, and OpenAI says that is the experience most people actually have.
That caveat matters, but it undercuts the defense for developers. API access is what you build on, and the study suggests the two leans come from different places. With search switched off in a 30-race sample, GPT still picked the Democrat in 24 races. Grok’s Republican lean vanished, and it refused to pick in 27. Researcher Tapan Parikh reads that as GPT’s lean being in the weights and Grok’s being in its search terms and sources.
There is also a detail you can’t audit. The study says every Grok API request carried about 1,240 tokens of extra input the researchers neither wrote nor could remove. Parikh calls them possible hidden instructions, though the evidence stops short of proving that. SpaceXAI didn’t comment. If you ship either model for ranking or advice, you inherit a lean you can only test for, not inspect.
2. The PHP embargo leaked, or someone got there first
The most unsettling detail in the PHP2Shell story is the timeline. Amazon says it saw exploitation about a week before the patches went public on Wednesday, with signs as early as October 17. The patches were meant to be held under embargo so big hosts could fix things first. Someone clearly knew before that.
TechCrunch lays out the possibilities: a leak, independent discovery, or espionage. The PHP Foundation blamed one company for releasing patches hours early, but that can’t explain a week of activity. Two unauthenticated remote-code-execution bugs in software behind roughly 70% of the web-facing internet is a bad place to rely on good-faith secrecy. The bugs were first found in 2024 and judged unexploitable, so they sat unpatched until researchers found a way to exploit them.
The embargo protected the giants. Banks, retailers and hospitals pulled services offline to patch. Smaller teams are the ones TechCrunch expects to be slow, and Palo Alto Networks and Trail of Bits have now confirmed real compromises. If you run a self-managed PHP stack, assume you were in the exposed group and check your logs back to October 17.
3. Cloudflare’s quantum fix for data sovereignty is a compromise
Cloudflare’s new post-quantum key manager exists partly because the fancier option wasn’t ready. The company says no production-ready post-quantum attribute-based encryption scheme exists, so it dropped the flexible 2022 design for a simpler one. Each region gets a symmetric root key, and service-specific keys are derived from it using hybrid ML-KEM-768 and X25519.
The threat is harvest now, decrypt later: stolen ciphertext sits in storage until quantum computers can crack it. That matters most for governments, banks and hospitals whose data stays sensitive for decades, which are also the customers with residency rules. The design fails closed. A server without the right regional key rejects the request, and keys live in the kernel keyring, so a restart wipes them.
The tradeoff is blast radius. Cloudflare says that compromising a regional root key could expose every service key derived from it for that epoch. The product is also in closed beta, with general availability planned for Q1 2027, and this is Cloudflare describing its own system. It is a credible example of PQ moving into production, but not yet proof that customers are rushing there.
4. Daytona’s $1B bet: agents need their own computers
Daytona raised $60 million at a $1 billion valuation just eight months after a $24 million Series A. FirstMark Growth led both rounds, and the total raised is now $91 million. The angel list reads like a founders’ reunion: Patrick Collison, Tobi Lütke, Aravind Srinivas, Eric Yuan, Olivier Pomel and others.
The pitch, from founder Ivan Burazin, is that agents don’t behave like long-lived services. They spin up thousands of short, stateful jobs nobody planned for, so each needs its own sandboxed computer, with Linux, Windows, macOS and GPUs on one platform.
The thesis is plausible, but the announcement offers no revenue, customer or usage numbers. A famous angel roster shows who wants a seat, not that the market exists. The harder question is whether agent sandboxes become a separate cloud category or a feature the big clouds absorb. The source doesn’t address that, and Daytona’s valuation is a bet on the first answer.
5. Pentagon: prediction market bets could tip off enemies
The Pentagon has apparently decided that a betting market can be an intelligence leak. Ars Technica’s headline says Hegseth warned that prediction market bets warp military judgment and endanger troops. Its excerpt says a Pentagon memo describes insider trading as giving US enemies key intel on military operations.
Only the headline and excerpt were available, so the memo’s specifics, including any proposed remedy, remain unknown. The argument is still clear. If someone with operational knowledge can profit by betting on an outcome, the price itself can reveal that knowledge to anyone watching.
That turns an insider-trading problem into an operational-security one. For prediction market builders, the open question is whether the Pentagon wants better controls or a ban on markets tied to military operations.