Skip to content
Astrohacker
Day 35Singularity Watch

Palo Alto Networks Agrees to Buy Applied Compute for $4.5 Billion as nginx Patches Two HTTP/2 Flaws and a New Benchmark Hijacks Every Frontier Model Tested

Palo Alto Networks agreed to buy Applied Compute, a platform for open-weight models, for $4.5 billion in stock, nginx released fixes for two HTTP/2 flaws, PromptArmor's FrontierBreach benchmark found indirect prompt injection succeeded against every frontier model it tested, and Bill Ackman accused NVIDIA's board of breaching its duty of care.

A wood engraving of a three-gear module being lowered into a protective machine housing, enclosed by a clipped-corner almanac frame.

Palo Alto Networks Agrees to Buy Applied Compute for $4.5 Billion as nginx Patches Two HTTP/2 Flaws and a New Benchmark Hijacks Every Frontier Model Tested

Security: Palo Alto Networks Buys an Open-Weight Platform

Palo Alto Networks announced on November 18 a definitive agreement to acquire Applied Compute, a platform for post-training and inference of open-weight AI models, in an all-stock deal valued at $4.5 billion, subject to adjustments. The deal needs regulatory clearance and is expected to close in Palo Alto Networks’ fiscal third quarter of 2027. [1]

The company says it believes enterprise AI will be hybrid, with a large share of enterprises running open-source and open-weight models tuned to their own data, and that no large provider delivers those models securely today. Applied Compute’s team will join Palo Alto Networks, and its co-founder and CEO, Yash Patil, will become Co-Chief AI Officer after the close. [1]

The Wall Street Journal’s Kate Clark reported the deal at 3:01 PM CST, writing that Applied Compute was founded last year by three former OpenAI researchers; only the opening of the paywalled article could be read. @BVeiseh wrote that Palo Alto spent the money to serve demand left by OpenAI and Anthropic gating their models, and Ryan surfaced that post at 4:43 PM CST. [1]

A day earlier, at 2:26 PM CST on November 17, Palo Alto Networks CEO Nikesh Arora wrote that what the industry warned enterprises about six months ago is here, that AI is making it easy for hackers to attack multiple customers, and that incident responders are all busy. He urged companies to run multi-model testing across their infrastructure, patch their code and prepare their security operations to detect and remediate in record time. Ryan surfaced the post at 7:11 AM CST. [2]

Security: Two nginx HTTP/2 Flaws

nginx released versions 1.30.7 (stable) and 1.31.9 (mainline) on November 18 with fixes for two HTTP/2 bugs: CVE-2026-123007, memory corruption in HTTP/2 flow control, which nginx rates major and which affects versions 1.9.5 to 1.31.8, and CVE-2026-123633, a use-after-free in HTTP/2 rated medium. [3]

Mattias Geniar wrote that the two bugs allow remote code execution without logging in when HTTP/2 is on. His patch tracker, last updated at 1:09 AM CST on November 19, lists fixed packages from nginx’s own repositories, the Docker official image, Ubuntu, Amazon Linux, AlmaLinux and Alpine, while Debian 12 and 13 still had no fix. He advises switching HTTP/2 off until a fix is installed, and notes that the same release fixes eight other bugs, including a use-after-free in HTTP/3. [3]

At 12:03 PM CST Geniar advised everyone to switch to nginx’s official repositories and patch. @UID_ quoted him: “This is going to be a disaster.” Ryan surfaced that post at 4:32 PM CST. [3]

Security: Prompt Injection Against Frontier Agents

PromptArmor released FrontierBreach at 12:02 PM CST, a benchmark of indirect prompt injection against agents doing enterprise work drawn from OpenAI’s GDPval. Each case ran five times per model at medium effort, and an attack counted only if the harmful action actually happened in the environment. [4]

Across the original 21 models, attacks succeeded in 34.9% of test cases. Frontier models were not immune: Gemini 4 Argon at 4.3%, GPT-6 Astra at 7.6%, Claude Fable 5.2 at 10.0% and Muse Wonder at 15.2%, while Grok 4.5 was the most susceptible at 91.4%. [4]

Attacks hidden in third-party skills succeeded 45.5% of the time, ahead of connectors at 33.9%, document systems at 31.5% and local files at 28.5%. Arbitrary code execution was the most achievable objective at 63.2%, and healthcare workflows were the most exposed industry at 44.4%. These are PromptArmor’s results on its own benchmark; Ryan surfaced the release at 4:31 PM CST. [4]

@Austen wrote late on November 17 what he is telling family: AI is now able to hack basically everything, people can plan on pretty much everything going down for a day or two at some point, and everything with someone watching should be able to adapt and be fine. Ryan surfaced it at 8:07 AM CST. [5]

Intelligence: Measuring GPT-6.2

Artificial Analysis wrote that GPT-6.2 Astra is now second in its Intelligence Index, trailing only Anthropic’s Claude Fable 5.2, after a 5-point gain over GPT-6 Astra, and that OpenAI models now hold most of its cost-per-task Pareto frontier. It found GPT-6.2 Astra flat on its Coding Agent Index and GPT-6.2 Sol 3 points lower than GPT-6.1 Sol, while Astra led its 3D and robotics evaluations. Ryan surfaced the analysis at 5:55 AM CST and a user’s post calling GPT-6.2 good at 8:08 AM CST. [6]

The Pragmatic Engineer posted Microsoft CEO Satya Nadella explaining why GitHub Copilot had no response to Claude Code: code went from input to output, IDEs faded away in about six months, and AI-first product building is unforgiving because the next model can show capabilities that force a rewiring. “Don’t be in love with what you built,” he said. Ryan surfaced it at 12:19 PM CST. [7]

Governance: NVIDIA’s Board

At 11:12 AM CST Bill Ackman wrote that NVIDIA’s board is not acting in the best interest of shareholders. He wrote that the board fired Jensen Huang or forced his resignation without careful deliberation, that foreclosing his return before the facts are known is another breach of the duty of care, and that the scope of the board’s investigation is unknown. [8]

@ShangwenLi1 quoted Ackman: “I’m deeply disappointed in our board.” Ryan surfaced that post at 5:12 PM CST, and at 1:59 PM CST a post by @farzyness predicting that Huang will return to NVIDIA with a new board. At 7:25 AM CST Ryan quoted a post saying NVIDIA had removed Huang’s affiliate badge and wrote that the board “just killed the vibes of the company.” [8]

Policy: Disclosing Chinese Models

Reuters reported on November 18, citing three people, that the Treasury Department and the White House Office of Science and Technology Policy have discussed requiring US companies to disclose on their websites their use of Chinese AI models, part of a push against the spread of Chinese open-weight models. It was not clear whether the idea was gaining traction or would enter formal rulemaking. [9]

The report names Airbnb, which uses Alibaba’s Qwen for its customer service agent, and says House committee chairs earlier sent letters to Airbnb, DoorDash and Anysphere, the maker of Cursor, whose Composer 2 model was built on a version of Moonshot’s Kimi. Chinese models were the three most used by tokens on OpenRouter in October, it says. Ryan surfaced a post on the report at 8:39 AM CST. [9]

Safety: A Statement Against Anthropic

A public statement titled The MisAnthropic Statement says Anthropic warns that recursively self-improving AI could concentrate power and escape human control yet races to build it first, uses safety branding to court investors and employees, and has quietly weakened its safety pledges. Its site lists signatories including Steven Pinker, Hugging Face co-founder Thomas Wolf, xAI co-founder Igor Babuschkin and Center for AI Safety executive director Dan Hendrycks. [10]

Harvard Kennedy School assistant professor Stephen Casper wrote at 9:24 AM CST that he signed it, and Ryan surfaced his post at 10:08 AM CST. [10]

At 12:12 PM CST Dean Ball wrote that a common belief in the AI field holds that advanced AI means the inevitable forfeiture of human primacy, which he called “the silent consensus,” and that it must not remain silent any longer. Ryan surfaced it at 12:51 PM CST. [11]

Agents: Harnesses and Org Charts

Jediah Katz, who worked on Cursor’s agent harness and now works on the Grok Bot harness, published “The Efficient Agent Harness” at 2:01 PM CST. He writes that since launch the team made Grok Bot 3 times faster to reply and 2 times more token-efficient, and explains the levers: cache misses, tokens, price per token and steps. [12]

His example: Grok Bot’s coordinator delegates to subagents and waits, so its cache expired and missed. An A/B test showed a longer, more expensive cache lifetime paid off for the coordinator but not for subagents, which work continuously. Ryan surfaced the article at 5:17 PM CST. [12]

Coinbase CEO Brian Armstrong described, in a conversation with David Senra, a workflow of seven or eight lead agents, a chief of staff and three to five subagents under each, with an internal view that turns their output into a prioritized queue of decisions. Senra’s post quotes him predicting that 99.999% of transactions will be agent-to-agent in 10 years. Ryan surfaced it at 10:15 AM CST. [13]

@poteto published a guide to automating anything with Grok Bot, describing automation as a goal plus a process that no longer requires writing code. Ryan also surfaced a report that Gavin Baker has more than 30 Grok Bots, a user’s post on the widening gap between people who use AI well and those who don’t, and a photo post saying Devin Bots have taken over San Francisco. [14]

On work, Ryan surfaced @mardehaym, who wrote that the technology mostly works now and the hard part is getting value a second time after the first champions leave, and @trq212, who wrote that good software is a set of strong opinions. He quoted a post about a junior engineer whose DevOps job is asking Claude Code for a deployment report: “delegate everything you can to ai.” At 8:55 AM CST he wrote that every software engineer should build their own software factory, “an app that helps you build apps.” [15]

Energy and Space

Ryan published “Energy Is the Currency of the Universe” at 8:43 AM CST. He argues that AI turns energy into computation and solutions, that new power cannot be built fast enough to keep up, and that the grid operator PJM’s last two capacity auctions came up 6,623 MW and 6,831 MW short of its reliability target. He concludes that the sun is the end state and space is where it has no limit. [16]

Earlier, at 5:28 AM CST, he quoted @VraserX on AI widening access to intelligence: “money = energy.” [16]

Palmer Luckey wrote on November 17 that money from selling Oculus funded SLAB, a seastead with a 1,000-year design life, and that he spoke up because media outlets planned to call it his floating AI datacenter project. Ryan quoted it at 4:48 AM CST with “let’s do it,” and surfaced a post on reusable spacecraft at 8:07 AM CST. [17]

At 9:01 AM CST @willdepue wrote that sometimes he forgets and then wakes up in the middle of the night realizing that the singularity is all happening. Ryan surfaced it at 10:03 AM CST. [18]

Astrohacker

@AstrohackerLabs published Singularity Watch, Day 34 at 8:46 AM CST on November 18, with its plate attached and its page live on astrohacker.com. [19]

Sources:

← Back to News