Sec & AI News — 14 November 2026

11 min readBy Nathan House
Get every new Sec & AI News issue
Straight to your inbox. No spam.

🔴 Jensen Huang Resigns From Nvidia Amid Spy Investigation

On 8 November, Jensen Huang emailed Nvidia staff: "I was unfaithful to my wife." The same email said the board had started an independent investigation into "national security concerns." The FT reported that US intelligence believes Alice Huang, an executive at Singapore customer Megaspeed, was recruited by China's Ministry of State Security. Megaspeed is suspected of diverting restricted Nvidia chips to China. It denies wrongdoing, and no charges have been filed. The board says it has "no reason to believe that either national security or NVIDIA secrets were compromised." His memo names no one. On 11 November, The New York Times reported a relationship with an executive who reports to him. The next day, Huang resigned as CEO and left the board, effective immediately. CFO Colette Kress is interim CEO. Nvidia reports its quarterly results on 17 November, with a CFO doing both jobs.

🧭 AI-Driven Security Engineering Jobs: This Week

Insider risk and export control are old jobs. This week they made the front page. My jobs archive now holds 989 listings from named employers in 33 countries, and AI as the tool is required on every one. What changes with seniority is what it asks for. At entry level, 55% of postings want you to build the tooling, not just use it. At staff level it's 64%. At leadership level, 73% ask you to lead its adoption. The job of the week sits on top of the Nvidia story. Microsoft's Principal Security Operations Engineer for Quantum Security pays $143K to $275K in Redmond. It owns external and insider threat incidents, and wants "familiarity with export control (EAR / ITAR) and government program environments." The AI Master's Program teaches AI-Driven Cyber Security Engineering. No coding background required. Application only.

📋 GRC Job Ads Now Name AI in 9 of 11 Postings. Five Skills Keep Coming Up.

On 12 November we pulled every live job from 66 company job boards on Greenhouse. That came to 10,380 jobs. "9 of the 11 live security GRC analyst ads in our November sample name AI or automation in the job's duties, and 7 name AI itself." Only 1 doesn't mention AI anywhere. The same request keeps coming up: "hand the evidence work to AI, then prove whether the AI got it right." Five skills keep appearing. Automate evidence collection. Prove the AI got it right. Govern the AI itself. Code enough to build small tools. Write controls as tests. The pay is real. Vercel advertised a GRC Analyst at $134,000 to $202,000. My new video covers how analysts move up by proving the AI wrong.

🔐 Anthropic and Google Both Keep Their Strongest Cyber Model for Vetted Defenders

Anthropic released Claude Fable 5.2 and Mythos 5.2 on 9 November. They are "the same underlying model." Fable is for everyone, and its safeguards "will block most cybersecurity-related tasks," including penetration testing and exploit generation. Blocked requests get an answer from Opus 4.8 instead. Mythos goes to organisations approved through the Cyber Verification Program. On Anthropic's internal test, it found 10.2% more vulnerabilities than Opus 5.5. A footnote says vulnerability identification is "roughly on par with Opus 5.5." Google made Gemini 4 Carbon public the next day, with the same split. Public Carbon has cyber safeguards. "Carbon for cyber" comes only through Google's Fairwind Program. Artificial Analysis scores Fable 5.2 at 65 and Carbon at 60. Fable costs $8 per million input tokens and $40 per million output. Carbon's $2 and $10 introductory price ends on 30 November. I tested where Fable 5.1 drew that line.

🦊 Arctic Wolf Links Open-Source AI Pentest Agent AntiFox to Real AWS Intrusions

Arctic Wolf Labs published on 10 November. AntiFox is "an open source, Chinese language agentic pentesting platform." Its README says it was built on ARTEX v0.1.1, the framework behind last week's WorkOS breach. In several intrusions, Alibaba Cloud systems hosting AntiFox were "associated with confirmed malicious actions." In one, the source cycled through multiple AWS identities. It made "hundreds of cross-service API requests within approximately half an hour." Most were denied. One wasn't: a "successful GetSecretValue request against a production secret." Arctic Wolf's advice belongs on a sticky note: "Do not assume a sequence was unsuccessful because most requests were denied." The link is through infrastructure. It doesn't prove an agent ran every step. Nothing was hijacked, either. The attackers ran their own copies. AWS told The Hacker News its services "were not compromised and operated as designed."

🛠️ A Detection Engineer Job Was Advertised at Up to $330,000, With No SOC Behind It

Fluidstack runs AI compute infrastructure. It advertised a Staff Detection Engineer to run incident response end to end "without a large SOC to fall back on." That's the top of the ladder. It asks for 8+ years, and the posting has since been republished at $224,000 to $344,000. Lumin Digital's ad, at $155,000 to $180,000, said the quiet part plainly: "This is not a traditional SOC analyst seat." AI-driven triage and SOAR platforms "now handle the bulk of routine alert processing." The Tier 1 work people learned on is going first, so the bottom rungs are breaking. The way up is building the detections the automation runs. Employers want "a detection in version control, tested against historical data before it goes live, with a false positive number you can say out loud." AntiFox, the item above, is what those detections are now up against.

🧠 NEAR AI Was Breached Through an SGLang Bug. Its Secure Enclave Didn't Help.

NEAR AI published an incident report on 9 November. The attacker "exploited a remote code execution vulnerability in SGLang, the open-source inference engine running on our servers." That gave code execution on one inference server, then access to wider infrastructure. They downloaded account details for "approximately 100 users." They also reached 18 agents on a legacy deployment outside the confidential computing environment. From those, they took chat history, settings and logs. Two Cloud customers had API keys exposed in logs. On 8 November, the attacker could also have watched live requests on that server, "including requests running in confidential mode." NEAR AI's own summary is the useful part. Confidential computing "does not protect against a vulnerability in the software running inside the environment." The report names no CVE and no SGLang version. NEAR AI says it reported the bug to the SGLang project.

🪶 Anthropic's Scanner Found a Critical Apache HTTP/2 Bug Eight Days Before Google

Apache HTTP Server 2.4.71 shipped on 11 November. It fixes CVE-2026-119618, rated critical: heap corruption and RCE in mod_http2. Versions 2.4.55 through 2.4.70 are affected. Wiz says a "remote, unauthenticated attacker" can get code execution. Apache credits two finders. Anthropic's OSS Scanner reported it on 26 October. Google Cloud Vulnerability Research reported it on 3 November. Wiz says 64% of cloud environments have vulnerable resources. Its own caveat: that figure "does not take configuration into account." HTTP/2 is off by default, so real exposure is lower. One of the two finders was a machine. When Wiz published, there was no public exploit and no reported exploitation.

🧰 N-able's N-central Was Exploited Hours After a Public Exploit, Patched Servers Included

On 8 November, researcher Sean Bales published proof-of-concept exploits for more than 30 alleged N-central flaws, without telling N-able first. N-able called it "non-responsible communication." It took hosted N-central offline on 9 November and told on-premises customers to do the same. Hotfix 2 landed on 10 November, fixing seven CVEs. Then a new chain appeared. The 4EverAble2Shell proof of concept went up at 03:58 UTC on 12 November. Huntback's decoys saw the first attempt at 08:30. By 12:50, a mining pool listed 21 rigs, each serving an N-central login page. "19 of the 21 report build 2026.4.1.26," which is Hotfix 2. "Applying Hotfix 2 did not protect them." Exploitation needs the Legacy API, which N-able says is off by default. N-able's advisory said it had seen no compromise.

🧊 France Blames Iran's IRGC for Sabotaging Fridges at 700 Supermarkets

In the night of 11 to 12 October, attackers hit the refrigeration controls of "more than 700 retail sites in France." This week France named the attacker. It says SHAHID KAVEH is "an offensive cyber unit affiliated with the Electronic Warfare and Cyber Defense Organization of the Islamic Revolutionary Guard Corps." The foreign ministry says the attacks caused "significant material damage, sparking fires and spoiling foodstuffs." It puts the losses at "tens of millions of euros." It calls this "the first offensive cyber campaign carried out purely for the purposes of sabotage on French soil." France will summon Iran's ambassador and propose further EU sanctions. BleepingComputer reports the attackers changed defrost settings to high temperatures.

🧾 CISA Lists the OpenSSH Bypass as Exploited, Plus an AI Model Server

CISA added CVE-2026-123158 to its exploited list on 9 November, with a 12 November deadline. It's the OpenSSH hole on musl-based Linux from last week's issue, and the first CVE number we've seen for it. The same day, CISA added GPUStack, a server for running AI models. Its flaw "could allow unauthenticated administrative access," plus secret disclosure and code execution. The cause is a trust shortcut. GPUStack treated requests from 127.0.0.1 as admin. Behind a reverse proxy on the same host, "every remote request arrives at GPUStack with TCP peer 127.0.0.1." "No header or credential is required." The advisory says it was exploited against a v2.1.1 deployment. The attackers created custom backends running reverse shells. The fix is version 2.2.0. On 10 November, CISA also added two chained Barracuda Email Security Gateway flaws.

🎙️ Claude Now Joins Your Video Calls and Opens Other Companies' Apps Inside Itself

Anthropic launched Claude Meetings in beta on 12 November. Claude "can now join your video calls as a visible participant." It saves "a transcript, summary, and screen descriptions, not audio." By default, it stays out of calls with outside guests, and leaves if one joins. Anthropic's own advice: "make sure everyone on your calls is OK with being transcribed." Two days earlier came Sheets and Whiteboards in beta, plus interactive connectors. Apps like Asana, Canva and Dropbox now open inside Claude, "signed into your own account." The marketplace lists more than 9,000 connectors and plugins, "including community-built ones." Your connector allow-list is now a security control. Enterprise admins: Claude Data turns on by default on 20 November. When the model vendor ships the spreadsheet, every SaaS feature gets cheaper to copy. I wrote a 10-question moat scorecard for security SaaS. What still protects you is "who recommends you, what you're trusted to do, the data only you see, and the approvals a newcomer has to wait for."

⚖️ Google and OpenAI Now Both Sell an AI That Only Makes Decisions

Google launched the Gemini Decisions API on 11 November. You send one input and up to 1,000 questions. It returns probabilities and confidence scores in 150 to 250 ms. Price: $0.10 per million input tokens, with no output tokens billed. Questions can be yes/no, multiple choice or an ordered rubric. Google lists policy checks and tool selection as uses. OpenAI's Decisions API works the same way, with predicate, choice and score questions, on gpt-6-luna only. Both are classifiers that return probabilities. A classifier that gates an agent's actions is only as good as its calibration, and its input can still carry a prompt injection. Someone still has to check the answers, which is now a large part of a GRC analyst's job.

🚀 Automattic Spun Out Spacefast to Host the Apps Your AI Builds

Spacefast is now "its own company," per Automattic's 9 November post. You bring your own AI, "ChatGPT, Claude, Muse, Grok Bot, OpenClaw, Hermes and more." That happens through an API, a CLI or an MCP server. Spacefast supplies "Hosting, your own domain, and a SQL database." It charges for infrastructure, not AI credits. There are "Approval checks for sensitive agent actions, such as rollbacks and DNS changes." Sharing is private by default. Login is the gap: "Built-in app authentication coming soon." Until then, an agent-built app gets whatever login the agent wrote. I wrote this in July, and it still holds: "The moment you stop reading, you've switched from agentic coding to vibe coding, whatever you call it."

🏪 OpenAI Lets Businesses Claim Their Profile Inside ChatGPT

Businesses in ChatGPT launched in beta on 10 November. Through Business Hub, a company can "claim their business, manage their information, and see insights." Once verified, it can update key details and add product feeds with pricing. Insights cover mentions, website clicks and discovery topics. OpenAI says it doesn't share users' conversations with businesses. US ecommerce and online-service firms can also launch a Business Agent. It can be the destination for ChatGPT Ads. The page doesn't explain how most businesses get verified. Shopify merchants connect their store. Enterprise customers and advertisers go through an OpenAI rep. For everyone else, it doesn't say. ChatGPT has 1.2 billion weekly users. If a verified profile is how they'll find you, claim yours before someone else tries.

📈 Anthropic Sets Its IPO Range at $70 to $78 a Share

Anthropic launched its IPO roadshow on 10 November. The offering is 825 million Class A shares, expected at $70 to $78 each. Anthropic is selling 801,351,354 of them. Existing stockholders sell the other 23,648,646. Underwriters can buy up to 123,750,000 more. It has applied to list on the Nasdaq Global Select Market as "ANT." Goldman Sachs and Morgan Stanley lead the deal. There's no pricing date yet. At the top of the range, 825 million shares comes to about $64 billion, by my arithmetic. It's still a proposed offering. The registration statement "has not yet become effective."

🖼️ MiniMax's New Open Image Model Is Off-Limits in the UK, Outputs Included

MiniMax released H3-Image on 10 November. It's a 33.1-billion-parameter open-weights model that makes 16-megapixel images, 4096 by 4096. MiniMax's own post claims "4K image generation in 10 seconds." It names no hardware. It's on MiniMax's API, MiniMax Code and Hugging Face. The licence is the part to read. Its excluded territories are "the European Union, the United Kingdom, the Republic of Korea and the United States of America." You may not use the model "or any of their Outputs" outside the permitted territory. Products must display "Powered by MiniMax-H3-Image". Firms with over $1 million in annual revenue, or offering hosted services, need separate permission. It's the same clause as MiniMax's H3 video model. Downloading a model is now a compliance decision.