OpenAI's AI agents found hundreds of ways to gain admin access to Linux, and the entire experiment cost roughly $1.2 million. The people responsible for fixing those vulnerabilities now have quite a reading list.
Welcome, humans!
Yesterday gave us a pretty revealing look at where AI is heading. OpenAI demonstrated how quickly an autonomous system can hunt for security flaws, Anthropic put Claude to work on a decades-old biology mystery, and Apple started getting stricter about how much employees spend on AI.
Meanwhile, AI companies spent millions trying to influence an election, and Meta went shopping for another coding startup. Apparently, even the companies building the future of software think they need more help building software.
Let's get into it.
🆕 NEW From The Neuron
- OpenAI's Pierre deal brings its AI coding fight to GitHub: OpenAI announced plans to acquire Pierre Computer, a developer-tools company whose code-navigation technology already powers parts of Codex. Axios reported that the deal advances OpenAI's ambitions to challenge GitHub as AI agents take on more software development work.
- NVIDIA put AI agents on the hunt for aliens: NVIDIA is helping SETI use AI agents to operate radio telescopes and analyze incoming signals. The difficult part is determining whether a mysterious transmission actually came from space.
- OpenAI faced years of security warnings: Former employees reportedly warned about weaknesses in the company's defenses. OpenAI is now working with outside hackers who previously penetrated systems used by Anthropic.
Around the Horn: Thursday, November 5, 2026
OpenAI's AI agents found 948 ways to gain root access to Linux
OpenAI built an autonomous security pipeline that could investigate software weaknesses, reproduce attacks, and propose fixes. When researchers pointed it at the Linux kernel, the software at the heart of countless computers and servers, the results were substantial.
The experiment identified 2,308 potential security flaws and successfully reproduced 1,724 of them. Its agents ultimately developed working exploits for 948 distinct local privilege-escalation vulnerabilities. These flaws could potentially allow someone who already has limited access to a machine to obtain full administrator privileges.
The entire run cost approximately $1.21 million in API credits, or around $1,279 per distinct vulnerability with a working exploit. The agents also generated candidate patches for human review.
To appreciate the scale, consider OpenAI's earlier experiment. In July, a system powered by GPT-5.6 Sol found and exploited 166 distinct vulnerabilities. The newer system produced 948 in its October run.
Finding security flaws is getting cheaper. Processing hundreds of credible reports might be the next expensive problem.
OpenAI said Linux maintainers had already patched 61 of the identified vulnerabilities upstream by the time its research concluded. Another 43 had patches in preparation, while 299 were being publicly reported or discussed.
The attacks were developed in a controlled research environment, and they required an existing foothold on the target computer. The proposed fixes also need expert review before deployment.
Still, the experiment demonstrated how AI could increase the volume of security vulnerabilities discovered in widely used software. Linux maintainers now face the challenge of assessing those findings quickly enough to get verified fixes into users' hands.
🏆 TOP 5 NEWS (Around the Horn)
- Anthropic used Claude to investigate mysterious cell particles, helping researchers determine the first three-dimensional structure of a vault particle found in prokaryotes, which include bacteria. Science reported that the findings suggest these unusual protein structures may have a much older evolutionary history than previously understood. Scientists discovered vault particles in 1986, but their purpose remains unknown.
- Apple reportedly cut engineers' daily AI budgets, reducing some employees' allowances from $300 to $200. Further reductions to $100 are under consideration, while some nonengineering teams could see limits fall to $25. Employees may request exceptions. Apple is encouraging cheaper models and on-device AI as it seeks to control internal spending.
- AI political groups spent millions influencing Congress, but several heavily supported candidates lost their midterm races. Groups backed by AI industry supporters and advocates of stricter safeguards both reported mixed results. Separately, Gizmodo examined the election's AI-generated advertising, citing research that identified hundreds of synthetic campaign ads and substantial spending behind them.
- Anthropic signed its first paid publisher licensing deal, according to Adweek, securing Associated Press election coverage for Claude just before the midterms. The agreement gave the chatbot access to timely, attributed election information. Anthropic had previously avoided the publisher licensing arrangements pursued by rivals such as OpenAI.
- Meta acquired AI coding startup Conductor, whose software allows developers to assign separate pieces of a coding project to multiple AI agents. TechCrunch confirmed that Conductor's cloud service will shut down on November 20, after which customer cloud data will begin being deleted. Founder Charlie Holtz hinted at a new product arriving in early 2027.
Honorable Mentions
- India's cybercrime agency warned financial institutions about AI attacks, saying criminals are using agents to compromise systems and conduct financially motivated attacks at machine speed. The advisory followed a similar warning from the country's central bank.
- Uber reported just four AI-agent ride bookings during its busiest week in October. Uber Eats recorded another 500 AI-agent orders, compared with approximately 320 million combined trips and orders that week. The bots may be coming for our errands, but they're apparently still figuring out how to hail a car.
- AI shopping agents could disrupt Amazon and Walmart advertising. Reuters examined how automated shopping could change product discovery and threaten retailers' advertising businesses. Emarketer estimates that Amazon and Walmart account for 88% of the projected $74.4 billion U.S. retail media market.
- OpenAI is already preparing for the 2028 election debate. Global affairs chief Chris Lehane proposed a new AI-era social contract focused on economic opportunity and worker protections. Critics questioned whether the proposals would adequately constrain AI companies or primarily encourage wider adoption.
- Developers launched an AI-assisted revolt against expensive SaaS. OccupySaaS aims to reproduce cloud services such as Auth0 and Salesforce through clean-room, open-source implementations. Its alternatives include ZeroAuthy and SalesCompel, though reproducing the reliability of established commercial platforms remains a challenge.
🍪 TOP TREATS TO TRY
- ChatGPT Images 2.6: OpenAI introduced animated GIF creation from text prompts or uploaded photos. Users can adjust animation speed and edit individual frames. The update also makes image generation up to 50% faster and is rolling out across ChatGPT and Codex on web, desktop and mobile.
- Ripple-65B Preview: Ricursive Intelligence released an open-weight 65-billion-parameter model for coding and agent tasks. In its launch announcement, the company claimed its model surpassed several much larger U.S. open-weight competitors on coding benchmarks. Weights are available under Apache 2.0.
- Kimi K3.1 Pro: Moonshot AI introduced a 2.8-trillion-parameter multimodal model with a million-token context window. It supports text and image inputs, plus agent tasks such as coding and financial research. The model activates approximately 104 billion parameters per request rather than the entire network.
- Tencent Hy4: Tencent released a 770-billion-parameter model capable of interpreting charts and documents while operating graphical interfaces or command-line tools. Access is available through Tencent products and third-party platforms, with limited free usage in WorkBuddy and CodeBuddy through November 15.
- Pangram Image 1.0: Pangram launched an image detector that identifies potentially AI-generated areas within pictures through a visual heatmap. Its faster Flash model focuses on AI video detection, including frame-by-frame analysis. The company reports high accuracy in its own evaluations.
- Claude Code reduced background agent status-update costs: Anthropic changed how background coding agents generate progress updates. Haiku 5.5 now writes updates from short notes instead of repeatedly reading the entire conversation, cutting the cost of individual status updates by approximately 99%. Anthropic says sessions with background agents may use roughly 5% less of a user's allowance.
- Grok Bot can now make phone calls: A demo showed Grok Bot contacting a service provider to cancel an internet subscription after the user approved the call. The feature moves the agent into phone-based customer service interactions, where users still need to authorize consequential actions.
- Notion promoted shared memory for AI agents: Notion described how ChatGPT, Claude and other assistants could use the same workspace information through its connections. The approach aims to let people retain project context when switching assistants. Some functionality remains subject to Notion's rollout and access requirements.
- ElevenAgents expanded its contact center integrations: ElevenLabs announced native connections with NiCE CXone, Genesys Cloud, Amazon Connect and Five9. Businesses can deploy voice agents into existing call systems, with the agents able to handle routine requests and transfer callers to human representatives when necessary.
- OpenAI Daybreak Red arrived in AWS GovCloud: Approved government customers can access GPT-5.6 Cyber through Amazon Bedrock in AWS GovCloud US-East. The service supports authorized vulnerability investigations and security testing. OpenAI approval and separate AWS provisioning are required.
- Floot connects app development to ChatGPT and Claude: Floot's connector allows users to request full-stack apps from their existing chatbot conversations. The company says it handles databases and authentication, plus hosting and deployment. Its product description also advertises publishing to iOS and Android through the same workflow.
- Wanaka Studio builds playable 3D games from prompts: The AI game-development platform offers generated 3D environments and a visual editor for revisions. Its advertised capabilities include automatic character rigging, multiplayer features and one-click publishing.
🏢 Big Tech & Major Companies
- Nothing teased a screenless AI device for December. CEO Carl Pei described it as a productivity tool that couldn't simply exist as an app, but he did not reveal its exact purpose or design.
- Cat Wu left Anthropic to join OpenAI. The former Claude Code product leader is moving to OpenAI's core products organization as competition for enterprise coding customers continues.
- AI demand is diverting semiconductor supply from other industries, according to Siltronic CEO Michael Heckmeier. Chip manufacturers are prioritizing lucrative data-center orders, potentially limiting supply available for smartphones and personal computers.
💼 AI Productivity, Labor & Economics
- Perplexity reported a 3.5-fold increase in ad revenue in a customer case study featuring Get Maine Lobster. The company says its Computer agent helped create Meta advertising campaigns and increased return on ad spend by 63%.
🏛️ AI Policy, Governance & Safety
- Music publishers suffered a setback in their Anthropic lawsuit. A federal judge dismissed a copyright infringement claim against CEO Dario Amodei, allowing the plaintiffs to amend, and questioned whether publishers had sufficiently supported allegations that Claude reproduced protected song lyrics.
🎙️ Interviews, Panels & Podcasts
- Hugging Face's policy chief discussed open-model trade-offs. Irene Solaiman joined TechCrunch's Equity podcast to discuss how startups should evaluate open AI models, including security risks and the costs of maintaining the surrounding software ecosystem.
Previous Around the Horn Digests
Catch up on everything you missed:
- Wednesday, November 4: OpenAI's security warnings, Google's school search restrictions and Apple's concerns about autonomous agents.
- Tuesday, November 3: Anthropic's self-improvement warning, hundreds of abusive deepfake apps and election misinformation.
- Monday, November 2: OpenAI's subscription changes, Meta's advertising growth and Apple's China AI plans.
- Friday, October 30: OpenAI's agent monitoring experiments, Reddit's traffic concerns and Mistral's trillion-parameter model.
- Thursday, October 29: Gemini 4 Carbon's cybersecurity capabilities, an AI medical-advice lawsuit and GitHub's AI code surge.
- Wednesday, October 28: Government AI terminology, OpenAI's reported TPU plans and Anthropic's government expansion.
- Tuesday, October 27: OpenAI's mathematics research, Anthropic security incidents and SpaceX's GPU financing.
That's a Wrap
We've covered another day's worth of AI security experiments, surprising scientific discoveries and new tools.
If you've made it this far, congratulations. You've read more about Linux security vulnerabilities today than most people will in their entire lives. The AI agents may still have you beat, though.
Get the condensed version of tomorrow's biggest stories by subscribing to The Neuron.
See you tomorrow.
P.S. Know someone who follows AI developments as closely as you do? Forward this roundup and tell them to subscribe here.
