OpenAI’s Bot Found 948 Ways to Root Linux -- AI Brief November 8

Good day %%first_name%%. OpenAI says an unreleased model it built a pipeline around turned a $1.2 million run into working root exploits for 948 Linux kernel bugs, and a Sunday brief that opens there also has an unreleased Anthropic model credited with a quantum proof, an outside auditor saying AI is speeding up Anthropic’s own research, OpenAI’s $11 billion at a $1.25 trillion price tag, and the Democratic bills coming for AI in January.

OpenAI’s Bot Found 948 Ways to Root Linux SecurityWeek

* What happened: OpenAI said on Wednesday that a pipeline built around an unreleased cyber model flagged 2,308 candidate bugs in the Linux kernel, reproduced 1,724 of them, and built working exploits for 948 flaws that would let an ordinary user on a machine take full control, SecurityWeek reports. The Hacker News says that took 30 hours in a mid-October run and cost $1,213,430 in API credits, about $1,279 per exploit.

* Why it matters: Linux runs a huge share of the world’s servers, and an exploit like this is what an attacker who already has a foothold uses to become the administrator. OpenAI’s own write-up says an unpatched system with a low-privileged login “should be considered functionally equivalent to root access for a well-resourced threat actor.” It also notes that at what Pwn2Own and TyphoonPWN have paid for kernel exploits, the 948 would be worth $18.9 million to $66.3 million.

* What everyone’s saying: The line from SecurityWeek and The Hacker News is that AI speeds up both sides: discovery and exploitation, but also fixes. OpenAI says that within 10 days of the scan, over 100 of the flaws had been patched or had staged patches from other contributors, with about 300 more under public discussion, and that it has already sent hundreds of reports and patches to maintainers.

* My read between the lines: The price tag is the story: $1,279 a bug turns exploit-writing from a craft into a line item. Two cautions. SecurityWeek says the agents produced 1,169 working exploits and that 948 counts the separate patches behind them, so the headline number depends on how you count. And on Google’s hardened kernelCTF target, the same pipeline managed 21 exploits in 112 hours for about $33,000. This is OpenAI’s account of its own unreleased model, and it is keeping the technical details private for safety.

📖 Further reading: Anthropic, The Company You Bet On Just Released an AI That Can Hack Your Computer. Here’s the Real Story. — my April read on Project Glasswing, Anthropic’s autonomous vulnerability hunter, and the four-step audit for what a hacking-grade AI means for your own workflows.

A thousand holes in the Linux kernel are a machine’s problem. The small-business version of an unpatched hole is the phone ringing while you’re up a ladder or under a sink. Marblism sells AI employees that work inside the tools you already use, and Rachel is the one who answers phone calls and books jobs. No prompting skills needed: describe your business and goals, and the employees check in when they need your approval. New users get 10% off for life through this link, and there’s a 7-day money-back guarantee. Meet Rachel →

An Unreleased Anthropic Model Cracks a Quantum Puzzle Business Insider

* What happened: Physical Superintelligence, a physics research lab known as PSI, said Friday that a research system running a non-public Anthropic model proved the Aaronson-Ambainis conjecture, a quantum-computing question open since 2008, Business Insider reports. PSI says the system had a complete proof on the third day of the run, and that Lean, a proof assistant that checks every step by computer, accepted it.

* Why it matters: The conjecture concerns when a quantum computer can be exponentially faster than an ordinary one. PSI says the result shows that exponential advantages like Shor’s factoring algorithm need special structure in their inputs, which Business Insider reads as quantum computers having limits that were long suspected, while they’d still be a major upgrade for some tasks.

* What everyone’s saying: Researchers who had worked on the problem are impressed. MIT and Stanford researchers who settled a special case of the problem in August (Guy Blanc, Jordan Docter, Carmen Strassle and Li-Yang Tan) said in PSI’s announcement: “This is a huge result,” adding that “everything is brought together in a rather magical way,” and they’re writing a short exposition of the proof.

* My read between the lines: Hold the confetti until the experts weigh in. MLQ notes that the paper is on GitHub rather than arXiv, hasn’t been peer reviewed, and that Scott Aaronson hadn’t posted an assessment as of November 7. Neither PSI nor Anthropic has named the model, and Anthropic didn’t respond to Business Insider. The computer checked the steps. A human still has to confirm it proved the right theorem.

<