948 ways to root: OpenAI's agents break Linux at scale, Perplexity hires out teammates, and a voice model fits in 32 MB
An OpenAI pipeline reproduced 1,724 Linux kernel bugs and wrote working root exploits for 948 of them, then drafted the patches. Agents also moved into team chats and into Cursor's deploy button. And at the small end, a speech model shrank to less than a few phone photos.

Key points
- OpenAI says an agent pipeline built on an unreleased cyber model reproduced 1,724 Linux kernel vulnerabilities and produced 948 working root exploits in one mid-October run, for $1.2 million in API costs. Patches go to maintainers after human review.
- Perplexity launched Teammates, always-on agents that each take a role in a workspace and report in their own thread or in Slack. They need the $200-a-month Max plan.
- Cursor's new Build Live Apps lets its agents set up a database, sign-in and hosting, then hand you a live link. It is on paid plans with cloud agents turned on.
- Conductor, an app for running several coding agents side by side, says its team is joining Meta. No new sign-ups; its cloud service shuts down on 20 November.
- OpenAI closed $11 billion, the first part of a round at a $1.25 trillion valuation, Bloomberg reports. Two weeks ago the talk was of up to $30 billion.
- Underdog showed Woof, a text-to-speech model of about 32 MB and 31 million parameters that is free to download and can run without internet.
Two weeks ago we wrote that code is now written, attacked and checked by machines. This week a machine did all three to the Linux kernel in one run. Agents also got jobs in team chats and a button that ships finished apps. The money kept coming, a little slower than planned.
Seven days of AI agents
OpenAI publishes its Linux pipeline: 1,724 bugs reproduced, 948 root exploits.
Conductor says its team is joining Meta and stops new sign-ups.
Perplexity launches Teammates on its $200-a-month Max plan.
Cursor lets its agents build, host and share live web apps.
OpenAI closes $11 billion at a $1.25 trillion valuation, Bloomberg reports.
Underdog demos Woof, a 32 MB speech model.
Source: OpenAI, Conductor, Perplexity, Cursor, Bloomberg, TechCrunch
A thousand paths to root
OpenAI pointed a chain of agents at the Linux kernel. Scout agents picked the parts an ordinary user can reach, about 7% of the code. Analyst agents read them and flagged 2,308 possible bugs in 12 hours. Triage agents then tried to trigger each one in an isolated machine.
They reproduced 1,724. For 948 of them the pipeline wrote a working exploit that turns an ordinary local account into root, on a stock desktop Linux distribution. It also wrote 291 proofs for bugs that leak memory. OpenAI says the agents used at least 10 exploitation techniques it had not seen before.
The run cost $1,213,430 in API credits, or $1,279 per root exploit. It used an upcoming cyber model that is not public. Each finding comes with a root-cause note, a draft patch and the exploit, and a human researcher reviews it before anything goes to the kernel maintainers. OpenAI says it started with the critical ones and has not said how many fixes have landed.
The blunt line is OpenAI’s own. For now, a low-privileged account on a system that is behind on patches “should be considered functionally equivalent to root access” for a well-funded attacker.
Why it matters to you: if you run Linux servers, laptops or Android devices, kernel updates are no longer the kind you can leave for later. Install them, and keep untrusted code in sandboxes.
Agents with job titles
Perplexity launched Teammates. Each one gets a role in a shared workspace, works around the clock and reports in its own thread or in Slack. They can run several jobs at once and pull in projects and memory from Perplexity’s other tools. The Verge compares them with OpenAI’s Dots and Anthropic’s Claude Duet. The price is the difference: Teammates need the $200-a-month Max plan.
Cursor went further down the line, from code to a running product. With Build Live Apps, its agents set up a Postgres database, sign-in and hosting, then hand over a live link. It is open on all paid plans with cloud agents turned on. Cursor runs security checks before an app goes live. Sharing sits behind Cursor sign-in, and public links are for individual plans only.
And one of the tools for running agents side by side changed hands. Conductor says its team is joining Meta. New sign-ups stopped on 4 November, the cloud version goes read-only on 13 November and shuts down on 20 November, and paying users get refunds. The desktop app gets fixes until early 2027. The team promises “something new” early next year.
Why it matters to you: if you build on Conductor Cloud, export your work before 13 November. If you are trying Cursor’s live apps, treat the link as a demo until you have checked who can sign in and what data it holds.
Disclosure: AI Beat uses Claude, among other tools, to help research and check its articles.
The money, a little slower
OpenAI closed $11 billion, the first part of a round at a $1.25 trillion valuation, Bloomberg reports, citing people familiar with the deal. The investors were not named. Two weeks ago reports spoke of up to $30 billion at the same valuation. The rest of the round is still to come.
Small is useful too
At the other end of the scale, the startup Underdog showed Woof, a text-to-speech model of about 32 MB. It has 31 million parameters, is free to download and can run on most devices without an internet connection. The demo is a small game in the browser. Underdog has not named a licence or the languages it supports, and TechCrunch notes that a model this small likely won’t replace the big ones.
Why it matters to you: a voice that fits on any phone and needs no network is a real building block for offline tools, from reading aids to apps for when the connection is gone.
Also today
- Claude Fable 5.2. Anthropic released a new top model on 9 November. Output costs $40 per million tokens, down from $50, and it more than doubled its predecessor’s score on the Remote Labor Index, a test of real freelance projects, to 41.3%. All figures are Anthropic’s own. Read our full story →
The pattern
The agents moved from helping to doing. They break the kernel and draft the fix. They hold a job title in a team chat. They press deploy. The gap that is left is review. In OpenAI’s pipeline a human still signs off every patch. Everywhere else, that step is up to you.
Sources
- OpenAI: A thousand fewer paths to root (4 November 2026)
- The Verge: Perplexity launches its own take on AI agents for work (5 November 2026)
- Cursor changelog: Build live apps (5 November 2026)
- Conductor: Conductor is joining Meta (4 November 2026)
- Reuters via GV Wire: OpenAI secures $11 billion funding at $1.25 trillion valuation, Bloomberg News reports (6 November 2026)
- Bloomberg: OpenAI completes $11 billion portion of funding (6 November 2026, paywalled)
- TechCrunch: Underdog shows off a 32MB speech model in new demo (6 November 2026)
Comments
What do you think? No account needed. Please be kind; comments that attack people are removed.
No comments yet. Be the first.