News Roundup is my personal digest with the main things from newsletters and sites I follow, saved here for my own reference and ease of access.
SSHole is now exploited in the wild: CVE-2026-123158 is in CISA’s KEV catalog, and attackers are getting root on Alpine/musl SSH hosts
BleepingComputer · SecurityWeek · Ars Technica · Alpine Linux advisory · CISA · Nov 6–14
Last week’s Alpine/musl OpenSSH bypass now has an ID, CVE-2026-123158, and it’s under active attack. CISA added it to its Known Exploited Vulnerabilities catalog on Nov 9. GreyNoise recorded the first successful exploitation on Nov 7 at 00:46 UTC, about 40 hours after the fix shipped, and logged 16 successful exploitation chains from six IPs: system and container recon, cloud-metadata queries, attempts to read password hashes and root SSH keys, and edits to authorized_keys for persistence. In one case an attacker got root and installed TAC Agent, a Python command-and-control implant, which couldn’t reach its server. GreyNoise also says a group it tracks as GT10 hit its Southeast Asia honeypots with a backdoor that has Chinese comments and looks AI-generated. watchTowr says the activity it has seen so far is related to botnets and cryptocurrency mining.
The bug hasn’t changed. On musl systems where sshd is built without PAM (Alpine’s default), a 257–1,024-byte password makes musl’s crypt() return *, which matches the * password field of key-only accounts such as root:*, key-only adduser users and the tiny-cloud alpine user. The attack needs at least one other account with a $5$/$6$ hash. Fix: OpenSSH 10.6p2 or Alpine’s patched packages (3.24 → 10.3_p1-r2, 3.23 → 10.2_p1-r1, 3.22 → 10.0_p1-r11, 3.21 → 9.9_p2-r1, edge 10.6_p2-r0), then restart sshd, because the package upgrade doesn’t replace the running daemon. Otherwise set PasswordAuthentication no. To look for compromise, check for unexpected successful password logins to * accounts, SSH keys you didn’t add, and TAC Agent files or startup services.
Node.js 26.13.1, 24.22.3 and 22.23.5 fix 16 CVEs, including an HTTPS bug that can leak other connections’ memory and TLS private keys
Node.js security blog · Nov 11
The November security releases shipped on Nov 11: v26.13.1, v24.22.3 and v22.23.5, with 16 CVEs (7 high, 6 medium, 3 low) and updated llhttp, undici and sqlite. The worst, CVE-2026-123864, affects every line. A zero-length TLS write re-enters llhttp and causes a use-after-free / out-of-bounds read that can leak other connections’ memory, including TLS private keys. It triggers when a server calls res.flushHeaders() before the body is done and then res.end() with a non-empty body.
The other highs include an unauthenticated HTTP/2 teardown assertion abort (CVE-2026-123863, 24/26), an HTTP/2 GOAWAY session leak (CVE-2026-123865), an HTTP/2 stream-teardown use-after-free that leaks response memory (CVE-2026-123866), a crash from an uncaught exception in deferred TLS handshakes (CVE-2026-126051, 24/26), a TLS 1.2 renegotiation servername change that bypasses per-host client-cert auth (CVE-2026-123872), and a respondWithFile use-after-free (CVE-2026-123874, 24/26). Mediums include a 1-byte altsvc overflow and an SNI mTLS fail-open. Several of the highs were credited to Chris Rohlf “in collaboration with Claude and Anthropic Research.” Anything serving HTTPS or HTTP/2 straight from Node, with no terminating proxy in front, should be upgraded and restarted.
The Fireworks AI breach spreads: a key Fireworks exposed led to Cursor’s fake Claude Code extension, and Notion AI chat transcripts leaked
Cursor · TechCrunch · Notion · ctxsec · Nov 9–13
On Nov 13 Cursor explained its October incident, and it started at Fireworks. An attacker obtained an access key Cursor had issued to Fireworks, which Fireworks exposed, and used it to copy a read-only bucket of coding-agent traces. They were mostly from 2025, came from sessions not in Privacy Mode, and were permitted for training. Some traces contained unredacted credentials: fewer than 0.01% of customers had API credentials exposed, and traces were exposed for fewer than 0.05% of users. With those credentials the attacker got into Cursor SaaS tools: archived non-production repos, support data and emails, usage metrics, agent-conversation summaries, and credentials users had sent to support. On Oct 29 a stolen credential let them plant tampered Anthropic.claude-code 2.1.312/2.1.313 in the cache behind Cursor’s extension marketplace (Open VSX itself was clean). It reached at most 5,000 users over about 11 hours. Cursor says the incident is contained (the last attacker activity, on Nov 1, was blocked) and that production, app releases, signing keys and customer repos weren’t touched. It moved all inference off Fireworks and rotated credentials across its internal systems.
Fireworks disclosed a second, worse breach. TechCrunch, citing Fireworks’ advisory, says an attacker got in through Fireworks’ cloud provider and reached a development system. Exposed: customer billing information and some datasets. Fireworks says it believes the attackers are financially motivated and focused on selling unauthorized LLM access, and ctxsec reports that some accounts also had secrets and incomplete model weights exposed. Notion confirmed that stolen data included Notion AI chat transcripts from its Insights feature, affecting around 1% of its customers. It disconnected from Fireworks, revoked exposed Notion API keys, and reported third-party keys found in messages to their issuers. If you used either product, rotate anything you pasted into chats or stored as a Fireworks secret.
OpenAI turns GPT-6 Cyber into products: Daybreak Red Team attacks your web apps, and Defense Factory drafts the patches
OpenAI · SecurityWeek · The Register · Nov 12
OpenAI’s Nov 12 cyber event produced two products. Daybreak Red Team (early access) points a coordinated team of agents at your web apps. They use source code and app context to find attack paths, check findings from scanners, bug bounties and disclosure reports, and re-test whether a patch actually stops the exploit. Isolated environments and “Guardian review” keep them in scope. Daybreak Defense Factory (a preview plugin in ChatGPT for Daybreak workspaces) traces vulnerabilities through your code and drafts fixes. It can import Wiz, Snyk and Endor Labs findings, scan connected GitHub repos and new commits on a schedule, and work with Codex on the patch. Engineers review the changes; nothing deploys automatically. The Register reports the agents run on GPT-6 Cyber.
Greg Brockman’s keynote line, via The Register and SecurityWeek: companies will be secure only to the extent of the frontier model they use to defend themselves. OpenAI is now selling AI pentesting and patching as products, not just a research model.
Open-weight models as attack surface: “dormant neuron” backdoors, and a free AI hacking framework used against 200+ orgs in 11 days
X · Dimitris Papailiopoulos · The Information · Jason Lemkin · Nov 10–12
Dimitris Papailiopoulos says every network has dormant neurons that can be rewritten to emit payload tokens into the residual stream. That’s enough, he argues, to backdoor an otherwise flawless agent model, and he says Opus 4.5 found such a backdoor back in March. In a follow-up he describes a technique he calls Lazarus: find a weak neuron, point its in-weights at a trigger, use SiLU to keep it quiet otherwise, and set the out-weights to scream a chosen token into the residual stream. If you pull weights from public hubs, where they came from now matters as much as the license.
Offense got cheaper the same week. On The Information’s “Information 50” call, as relayed in TI’s public post, Tenzai CTO Ariel Zeitlin said freely downloadable Chinese open-weight models are now good enough to break into corporate networks, and CrowdStrike said a China-linked group used ARTEX, a free AI hacking framework posted to GitHub in July, against more than 200 organizations in 11 days. Separately, Jason Lemkin said an AI pen-testing tool breached 200+ organizations in 60 days.
Jensen Huang resigns from Nvidia after a probe into his ties to Alice Huang; CFO Colette Kress is interim CEO and Congress wants him under oath
NVIDIA · SEC 8-K · CNBC · CBS News · Nov 11–14
Nvidia’s board announced on Thursday, Nov 12, that Jensen Huang has resigned as president, CEO and a director, effective immediately. The company’s 8-K says the board accepted that resignation on Nov 11. CFO Colette Kress is interim CEO and keeps the CFO job while the board searches for a permanent successor, considering internal and external candidates. The same announcement named director Harvey C. Jones as Nvidia’s first-ever chair of the board. Kress’s first public test is Q3 earnings on Tuesday, Nov 17, still on the original schedule.
The resignation followed public reports of Huang’s relationship with Alice Huang, a former executive at Singapore-based Megaspeed, and of a US investigation into whether that relationship and Megaspeed’s business posed a national-security risk. CNBC and other outlets have described Alice Huang as a suspected or alleged Chinese intelligence operative; that allegation has not been established in the company’s own announcement, which does not mention her. In Washington, House chairs Brian Mast, John Moolenaar and Rick Crawford asked Huang for a closed-door interview on Dec 1 and said they would subpoena him if he refused; Mast said Huang indicated he is willing to appear. Senate Banking chair Tim Scott has asked Huang and Commerce Secretary Howard Lutnick to appear in December.
A router at one remote FAA site grounded every US flight for about 4.5 hours; Microsoft 365 and Outlook were down most of a day
Axios · Reuters · US DOT · FAA · X · Gergely Orosz · Nov 12–14
On Friday Nov 13 the FAA issued a nationwide ground stop after communication failures across air-traffic control centers. Reuters says the trouble began at 3:36pm ET and that Transportation Secretary Duffy said airspace reopened at 8:15pm ET, a gap of about 4.5 hours, and called it the longest shutdown since 9/11. By 11:30pm, Flightradar24 had counted nearly 10,000 cancellations. On Saturday Duffy said a contractor, L3Harris, was replacing an old copper line with fiber and installing a new router at a remote site when the outage began; it’s not yet clear whether the cause was installer error or faulty equipment. Homeland Security said there was no evidence of a cyberattack. The FAA was back at full operations Saturday, with thousands of flights still canceled.
A day earlier, starting on Nov 12, Outlook and Microsoft 365 went down worldwide for many hours. Gergely Orosz, who was at Microsoft that day, said Outlook.com was down globally and that Microsoft employees could not send or receive email. After about 20 hours he wrote that the outage was being mitigated, that people at Microsoft told him it was a networking problem rather than DNS, and that the fix involved moving workloads from the legacy Exchange system to Azure.
Token budgets are now a hiring perk, while Booz Allen moves production off Anthropic to Nvidia’s open Nemotron
The Information · Applied AI · Nov 10–11
The Information reports that job candidates are negotiating higher or unlimited token budgets into offers, while Apple, Microsoft and Disney have tightened AI spend caps. Separately, Booz Allen CTO Bill Vass told TI the firm moved the models that run its cybersecurity and defense software from Anthropic to Nvidia’s open Nemotron — “the opposite of tokenmaxxing” — while developers still mostly use Claude Code to write software. Kilo Code told TI that Nemotron handled nearly 17% of the roughly 50 trillion tokens that passed through its platform in October.
HubSpot co-founder Brian Halligan returns as CEO to rebuild the company around a “customer relationship harness”
HubSpot · Nov 12
HubSpot said on Nov 12 that co-founder Brian Halligan is back as CEO, effective immediately, and that Yamini Rangan is stepping down as CEO and from the board. Rangan, who has led the company since September 2021, will stay through Dec 1 to help with the transition. Halligan, who ran HubSpot for 16 years through its IPO, is framing the move as a “refounding”: after inbound marketing and the CRM platform, he says chapter three is building a “customer relationship harness.” The company also named Adobe finance executive Karen Burgess as its next CFO, effective Feb 11, 2027, and reaffirmed its 2026 guidance.
Claude Fable 5.2 and Mythos 5.2: one model with two access tiers, lower prices, and a 30-hour formalization of Fermat’s Last Theorem
Anthropic · Simon Willison · Nov 9–10
This is Anthropic’s first frontier release since Dario Amodei’s September call to slow down. Claude Fable 5.2 and Mythos 5.2 are the same underlying model. Fable is generally available, including on Amazon Web Services, Google Cloud and Microsoft Azure. Mythos is available through Anthropic’s Cyber and Life Sciences Verification Programs. Anthropic says the models generate output up to 30% faster than 5.1 and cost $8/$40 per million tokens. An optional fast mode runs up to 2.5× faster for $16/$80. Willison’s llm-anthropic 0.31 supports the release, and he notes that Fable 5.2 is 80× the token price of Claude Haiku 5.5 or GPT-6 Luna.
Anthropic’s showcase includes “infeasible” migrations, devices designed from scratch, progress on open physics problems, and Fermat’s Last Theorem formalized in 30 hours. One practical trick from the prompting guide, highlighted by Willison: telling the model “You are the owner and long-term maintainer of this codebase” raised duplicate-code consolidation from under 10% to about 75%.
Anthropic’s IPO is set: ANT on Nasdaq at $70–78 (about $1.4T), pricing Nov 18 and trading Nov 19, plus a $26B placement with chipmakers
Anthropic · CNBC · TechCrunch · Yahoo Finance · Simon Willison · Nov 10–14
The roadshow launched Nov 10. Anthropic is offering 825 million Class A shares (801.4 million from the company, 23.6 million from existing holders) at $70–78, listing on Nasdaq as ANT, with a 123.75 million-share greenshoe. At the top of the range, TechCrunch and CNBC put the fully diluted value at about $1.404–$1.405 trillion, up from $965 billion in May but below the ~$2 trillion talked about a few months ago. Yahoo Finance reported on Nov 12 that the company said it will price on Wednesday, Nov 18, and begin trading Thursday, Nov 19. A separate $26 billion private placement of non-voting Class N shares at the IPO price goes to AMD, Broadcom, Google, Micron, Microsoft, Nvidia and TSMC, with Google and Nvidia at $10 billion each, per the filing as reported by TechCrunch. The founders’ Class F stock holds 50.1% of the vote, and the Long-Term Benefit Trust elects a majority of the board. Each of the seven named co-founders is selling about 3.4 million shares.
From the amended filing, as reported by CNBC, TechCrunch and Yahoo Finance: nine-month revenue of $33.2 billion (up about 1,200%), a Q3 operating loss of $536 million, an $81 billion October revenue run rate, and roughly $895 billion of committed spend. Willison’s notes on the roadshow video: Mythos got Claude running on AMD in three days with roughly 100% of kernels written by the model, and Anthropic now trains and serves across Trainium, TPUs, Nvidia and AMD.
Anthropic ships an office suite inside Claude: Meetings, Whiteboards, Sheets, Data and self-hosted Managed Agents
Claude · TechCrunch · The Verge · Nov 10
At a Nov 10 event Anthropic launched apps that run inside Claude. Claude Whiteboards and Claude Sheets are in beta. Whiteboards turn a plan into a visual timeline; Sheets builds trackers and budgets and exports to .xlsx or CSV. Claude Data, the former Dashboards product, is generally available and connects to warehouses and CRMs. Interactive connectors open third-party apps, including HubSpot, inside Claude. Claude Meetings, in beta on Team and Enterprise plans and starting with Google Meet, joins calls and turns the discussion into docs and tasks.
For engineering teams, Anthropic also introduced self-hosted Claude Managed Agents in private beta. It runs the agent harness, credentials and session history in your own cloud or on-prem, while model requests still go to Anthropic or the provider you choose. This is not an on-prem model release.
The AI Futures Project now expects a “superhuman AI researcher” in 2027, and Anthropic’s filing says its doubling threshold is met
AI Futures Project · The Next Web · Nov 10–12
Citing METR’s recent estimate that Anthropic is progressing about twice as fast as its historical rate, the AI Futures Project moved its forecast forward: all of its authors now expect a “superhuman AI researcher” — an AI that matches the best human researchers’ taste and can fully automate coding — at some point in 2027, assuming no substantial deliberate slowdown. “We’re at the beginning of the takeoff period” is the tenor of the update; they also argue for an industry-wide pause.
The same week, Anthropic’s Nov 10 amended S-1 made a related call. As The Next Web reports, the company said it meets its own Responsible Scaling Policy threshold for a dramatic, AI-driven acceleration — double the rate of progress without significant AI help — based on “what we expect from future models and our overall rate of progress,” not on what it has observed from current models. Its Mythos 5.2 system card had said that model does not itself double the internal rate of research progress.
OpenAI’s Q3: about $10B in revenue but a $10.9B operating loss, and an $11B first close at $1.25T
The Information · TechCrunch · Nov 10–12
The Information’s public recap says OpenAI’s third-quarter revenue was about $10 billion, with an operating loss of $10.9 billion, while Anthropic’s Q3 revenue neared $17 billion and its operating loss was $567 million. TechCrunch, in its IPO coverage, reported that OpenAI was valued at $1.25 trillion last week after raising $11 billion, nearly half of a $20 billion target. Both companies are heading toward public markets.
OpenAI’s product week: businesses get storefronts and agents inside ChatGPT, and a 35-second kill switch for research jobs
OpenAI · Nov 9–10
Businesses in ChatGPT is in beta. Companies can claim a profile in Business Hub, add product feeds with prices and availability, and see aggregated analytics on how they show up in ChatGPT. Eligible US ecommerce and online-service businesses can also deploy a Business Agent, a chatbot that lets people compare products and ask questions without leaving ChatGPT, and Shopify merchants can connect a store from Business Hub or the Shopify admin. OpenAI says 1.2 billion people use ChatGPT each week.
On safety operations, OpenAI described Control Room, an internal system for tracking research inference workloads and stopping them quickly. It covers agentic sampling for training and evaluation, data processing and more, and “halts model output within 35 seconds of a decision to intervene.” It’s an internal control, not an API feature.
Google ships Gemini 4 Hydrogen as a default everyday model, makes Carbon generally available (its price doubles Dec 1), and adds a Decisions API
Google · Simon Willison · X · Philipp Schmid · Demis Hassabis · Nov 10–13
Gemini 4 Hydrogen (Nov 12) is Google’s fast model for everyday coding and agent work. It scores 51 on the Artificial Analysis Intelligence Index at its default high thinking level, beating Argon’s 50, and 54 at xhigh. Pricing is flat across the 1-million-token context at $1.50/$7.50 per million tokens. Google is putting it in AI Studio, the API, the Gemini app, Google Search, and Antigravity, so it’s the Gemini most people will see. Gemini 4 Carbon, previously limited at launch, is now available more broadly. Willison records Carbon and Argon at a promotional $2/$10 through Nov 30, doubling to $4/$20 on Dec 1. Philipp Schmid also noted that Carbon is #1 on CursorBench and available in Cursor.
The new Gemini Decisions API (”Gemini Decision 1”) puts Google into the decision-model category alongside Jev and OpenAI. It answers yes/no, multiple-choice and rubric-scoring questions for $0.10 per million input tokens with free output, and unlike Jev it takes text, images, audio, video and PDFs. Willison shipped llm-gemini support for the releases and called the family’s pricing confusing; Hydrogen’s pelicans, he wrote, were notably worse than its pricier siblings.
Meta’s Muse Ember: a $0.10/$0.50 agent-workhorse model on Artificial Analysis’s price/performance frontier
Simon Willison · Meta · Nov 10
Meta released Muse Ember, its “smallest, most efficient model,” aimed at repetitive agent work: tool calls, coding subtasks, routing, extraction and parsing long documents. It costs $0.10/$0.50 per million tokens up to 200K context and $0.30/$1.50 above that, has six reasoning levels, and sits on Artificial Analysis’s price/performance (Pareto) frontier. Short-prompt pricing matches GPT-6 Luna and Claude Haiku 5.5; with the long-context tiers, Willison places it between them. His pelican tests across the reasoning levels were uneven, and the “high” setting failed to return a full SVG.
OpenRouter is now part of Stripe
X · OpenRouter · Nov 10
OpenRouter said on Nov 10 that it is officially part of Stripe, confirming an acquisition Stripe agreed to in August. The company says more than 10 million developers and companies use the gateway, and that joining Stripe will help it build faster while keeping the same neutrality pledge: models and providers still compete on quality, reliability and price. Neither company published a purchase price in the close announcement.
llama.cpp hits v1.0.0: a stable API with semver, FP8 KV caches, an MoE expert cache and a hardened llama-server
ggml-org / Hugging Face · GitHub release · X · Clem Delangue · Nov 13
After 3.5 years, llama.cpp and its tensor library ggml shipped v1.0.0 together. From now on, any breaking change to the public C API or ABI in include/ means a major-version bump; features, backends and model support are minor bumps; bug fixes are patches. The release removes 35 deprecated functions and adds OCP FP8 KV caches, a host-memory MoE expert cache, and support for 150+ architecture variants, including Kimi-K3.1-Pro.
llama-server gets hardening against SSRF, CORS and DNS rebinding, plus MCP tool cancellation, an OpenAI-compatible /v1/audio/speech endpoint and a redesigned web UI. Anything that embeds llama.cpp directly can now pin to a stable API, but moving to 1.0 means dealing with the removed functions.
AWS open-sources aws-emulator, an official local AWS clone in Rust built mostly by an agent swarm (29,684 commits in three weeks)
AWS · All Things Distributed · Simon Willison · Nov 11
AWS released an official local AWS emulator, written in Rust, so tests can run against a local copy of many AWS services without a paid account or API keys. It needs Docker on macOS and Windows. Willison picked out how it was built: an agent swarm made 29,684 commits in the first three weeks. Swami Sivasubramanian’s guest post on Werner Vogels’ blog tells how it came about and what the engineer’s job becomes when agents write the code.
Bun’s ahead-of-time compiler beats V8 by 1.5× on JetStream 3, but Bun v1.5 still hasn’t shipped
X · Jarred Sumner · Nov 9–13
Jarred Sumner posted a JetStream 3 result for the AOT compiler coming in Bun v1.5: 1.5× faster than V8 and 1.4× faster than JavaScriptCore, on untyped JavaScript. In a follow-up he also compared it against Static Hermes and QuickJS. Earlier in the week, still inside the Nov 9–13 window, he said OpenCode uses 41% less memory under the AOT build and OpenClaw starts 3× faster, and that reaching Rust-level performance may need new TypeScript or JavaScript syntax. The latest Bun release is still v1.4.4 (Oct 22), so there’s no ship date for v1.5 yet. (The JetStream post went up at 4:47pm PT on Nov 13 and was checked directly.)
SF Ruby Conf: RubyLLM 2.2 ships on stage with sandboxes, skills and Active Storage memory; Solid Queue gets a fiber mode
X · Carmine Paolino · This Week in Rails · Nov 10–13
Carmine Paolino released RubyLLM 2.2 on stage at SF Ruby Conf and wrote up how he tuned it for coding agents. The release adds coding sandboxes, skills support, memory and Active Storage file stores, and tools that can ask a person a question and wait for the answer. A fiber execution mode for Solid Queue was also shown at the conference. Rails core had a quiet week with no release (latest is still 8.0.5.2 / 8.1.4.1 from Nov 4). This Week in Rails (Nov 13) notes a Hotwire guide in Edge Guides and a deprecation of cross-table unscope(where:) behavior ahead of Rails 9.0. It also covers two quiet leak fixes: Action Cable error logs no longer include filtered parameters, and ActiveSupport::EncryptedFile#inspect no longer shows the key, plus a fix for :integer attributes truncating 64-bit strings.
Social Media
DHH: “back up open-weight models,” Omarchy 4.6 shares local inference across a tailnet, and Nvidia would be wrong to fire Jensen
X · DHH · Nov 11–13
After Nvidia announced Jensen Huang’s resignation, DHH wrote: “You might want to make sure you have a few open-weight models backed up and some basic gear to run them. Who knows what really happened at NVIDIA, but the world just lost one of free AI’s biggest proponents.” Earlier that day, before the announcement, he had argued that “NVIDIA would be wrong to fire Jensen” and that the matter was one for a board reprimand, not an ouster. He is building local AI into Omarchy: version 4.6, he said on Nov 11, lets you share any model running on an Omarchy machine with another machine on your tailnet, so a DGX Spark can serve an XPS 13. On Nov 13 he said he was putting finishing touches on an out-of-the-box local AI setup for RTX machines, “with the latest big jump in open-weight models.”



