Weekly · Shrinking Vulnerability Window and AI Consolidation · November 1–7, 2026
Key points
- The median time from vulnerability disclosure to exploitation has shrunk from a year in 2021 to one day in 2026, with nearly 80% of attacks occurring before official public disclosure.
- Marvell announced the acquisition of Lumentum for $115 billion to transition AI infrastructure to optical connections, triggering a drop in both companies' stocks due to regulatory risks.
- Saudi Aramco acquired a 49% stake in state-owned AI company Humain, investing $2.6 billion and linking oil revenues to a sovereign AI infrastructure project.
- OpenAI merged ChatGPT and Codex usage limits into a shared token pool and quietly enabled model training on Codex data, changing privacy terms for enterprise clients.
- Intel lost $140 billion in market capitalization after an earnings report where weak PC market forecasts outweighed strong data center performance.
Threats to AI Infrastructure: From Trend to Active Attacks
On November 1, an analysis by A New Stack, citing the J.P. Morgan Patchmageddon report, recorded the shrinking median time from vulnerability disclosure to first exploitation: from approximately one year in 2021 to one day in 2026, with a forecast of one minute by 2027. Nearly 80% of exploitations now occur on the day of or before public disclosure. Anthropic and Project Glasswing partners discovered more than 10,000 new high- and critical-severity zero-day vulnerabilities in the first month.
On November 2, Wiz described a wave of cyberattacks in October against at least nine AI platforms, calling it "Black October." This included the GhostAction campaign: one documented attack on October 8 affected 346 repositories, while the broader campaign impacted thousands. A malicious Claude Code extension was available on the Cursor marketplace for about 11 hours, affecting up to 5,000 users. A leak in OpenCode affected approximately 1.4% of Console users and was linked to a large-scale model distillation attack.
On November 6, Anthropic introduced a built-in privacy module, cc-plugin-privacy, in Claude Code v2.1.317 to block the sending of email addresses in User-Agent headers during web searches, after the issue was reported on November 5. On November 7, research by Vals showed that the Chinese open-weight model DeepSeek completed 73% of simulated network hacking tasks, nearly matching the closed Daybreak model from OpenAI (81%), at a token cost of $2–$3 versus $17–$46 for GPT-6-Sol.
AI Infrastructure Consolidation and Optics
On November 2, Marvell Technology announced the acquisition of optical component manufacturer Lumentum in a deal with a fixed stock exchange ratio, with an implied enterprise value of $115 billion. Lumentum shares will be exchanged for 4.5 Marvell shares. Marvell CEO Matt Murphy stated that AI infrastructure is undergoing a fundamental architectural shift toward optical connections. The deal is expected to close in the first half of 2028.
On November 3, the deal valued Lumentum shares at $1,132 each, which was 23.9% higher than the price on October 28. However, Lumentum shares fell by 18.6% and closed at $684.40, becoming the worst performer in the S&P 500 that day, while Marvell shares dropped by 12.7%, marking its worst three-day performance since October 2000. Cantor Fitzgerald analyst C.J. Muse named China's approval as the key question, and Counterpoint Research noted that the FCC reportedly considers restrictions on the import of Chinese optical transceivers.
On November 4, Marvell's bet on photonics was described as a response to explosive demand for optics from Nvidia and other companies, positioning Marvell at the intersection of digital signal processing and laser technologies for next-generation AI data centers.
Sovereign Capital and AI in the Gulf Region
On November 1, Saudi Aramco announced the acquisition of a 49% stake in state-owned AI company Humain, investing approximately 9.9 billion Saudi riyals ($2.6 billion) in cash, as well as assets in AI, computing, and digital technologies through a subsidiary. The Public Investment Fund will retain the remaining 51%.
Humain, launched in May 2025, plans to deploy three gigawatts of AI data center capacity by 2030 and has developed an LLM in Arabic. Aramco CEO Amin H. Nasser stated that AI helped the company achieve $11.3 billion in technological value since 2023. Aramco had planned a significant minority stake in Humain for at least a year.
OpenAI Policy Changes and Data Risks
On November 1, Parameter Labs disclosed a sandbox escape vulnerability in ChatGPT's HTML preview, which allowed the extraction of private user data via the virtual keyboard of the MathLive JavaScript library and the YouTube oEmbed endpoint. OpenAI closed the vulnerability and paid a $5,000 bug bounty. It was also announced that starting in December, usage limits for Plus, Pro, and Business plans would be merged into a shared token pool between ChatGPT and Codex.
On November 2, OpenAI confirmed in a help article that chats in ChatGPT and tasks in Codex would use a shared token pool from December, with Instant chat exempted. It was noted that the planned reduction of Pro 200 limits on October 29 was canceled, while limits for Pro 100, Pro 500, and Business Premium were doubled.
On November 3, users reported the quiet re-enabling of model training in Codex settings without explicit notification. A letter from the Daybreak Blue Trusted Access program stated that from November 29, OpenAI may use content from Daybreak Blue sessions on personal accounts to improve models, even when the "improve model for everyone" setting is turned off.
On November 5, OpenAI launched Images 2.6 in ChatGPT, Codex, and the API, generating images 50% faster than the previous version (Flare at low quality in 6 seconds), adding GIF generation, and noting that people create 4 billion images per week.
Chip Market: PC Weakness vs. AI Demand
On November 2, Intel shares fell more than 28% on Wall Street after the company released a fourth-quarter sales forecast below expectations, citing a slowdown in the PC market. On November 3, the earnings report showed a 29% drop in shares, erasing $140 billion in market capitalization, with the forecast for the December quarter described as below expectations partly due to weakness in consumer PCs.
Open-Weight Models and Inference Growth
On November 1, Confluent Vice President Sean Falconer published an analysis asserting that open-weight models are closing the performance gap with proprietary ones, citing data from Andreessen Horowitz: 37% of enterprises run five or more models in production. He recommended that most enterprises use inference providers such as Fireworks, Together AI, or Base10 rather than self-hosting.
On November 2, DeepSeek open-sourced DeepInfer — a framework for more efficient AI model deployment and reduced computational costs. On November 4, Tencent announced its LLM Hy4, stating that it enters the top tier of open-source AI models for the first time and is competitive with releases from Anthropic and OpenAI. An Interconnects study noted that Fireworks serves more than 40 trillion tokens per day for clients' custom models; this vertical has grown more than 2x compared to last summer.
What This Means
The shrinking vulnerability window to one day and the demonstration of attacking capabilities by open-weight models at $2–$3 mean that control over the AI supply chain and credential security must be restructured on a daily cycle. If you have an impact on AI development infrastructure, the risk of compromise through supply-chain attacks is increasing — heightened vigilance toward updates and environment isolation is required.
The $115 billion Marvell-Lumentum deal signals a shift in AI data center architecture toward optical connections. If regulatory approvals (especially in China) are delayed, the implementation timelines for new interconnect standards may shift — readiness for alternative communication architectures becomes more important.
Aramco's investment in Humain shows the transition of sovereign capital from passive investments to active co-ownership of AI infrastructure. If you work with Middle East markets, competition for computing power and data may intensify — it is worth considering new local providers in deployment strategies.
OpenAI's merging of the token pool and the quiet enabling of training on Codex data change cost and privacy calculations for enterprise users. If your obligations are sensitive to data processing terms, readiness to review contracts and data policies before the December launch is appropriate.
Intel's 29% stock drop shows that PC demand weakness can outweigh AI segment growth. If your supply chains depend on chip production volumes for client devices, readiness for price and delivery schedule volatility is justified.
What Remained in the Shadows
At first glance, individual events — the report on Postman's work with Claude Code and Codex, the launch of vSphere Express, and StateBench and Hightouch benchmarks — seem secondary. However, together they paint a picture of fragmentation in the AI tooling ecosystem: automatic configuration changes by agents without explicit consent, VMware's return to the SMB segment with a fixed price of $9,600, and evidence that hybrid models can achieve frontier-model accuracy at a 4x cost reduction. This indicates that control over configuration and inference costs is shifting from large providers to hybrid and local solutions, which requires continuous monitoring of changes in development tool settings.
Our read
The period shows that AI capabilities still outpace the means of testing and containing them: the vulnerability window has shrunk to one day, and open-weight models demonstrate attacking capabilities at $2–$3. The strongest counter-evidence is Anthropic's release of a privacy module and the closure of the ChatGPT vulnerability, showing that providers are responding to risks. Observed reversal indicator: a public announcement by a major financial institution about the deployment of open-weight models in its security stack.
Access to models and compute is concentrating: the $115 billion Marvell-Lumentum deal and Aramco's stake in Humain point to infrastructure consolidation. Counter-evidence — the growth of independent inference providers (Fireworks, 40 trillion tokens/day) and open-source releases from DeepSeek and Tencent. Reversal indicator: a hyperscaler acquiring one of the top-3 startups in the inference field.
Regulation is lagging behind the pace of deployment: OpenAI's quiet enabling of training on Codex data and the lack of explicit consent for Postman configuration changes show that rules are not keeping up with product changes. Counter-evidence — India's 6-hour reporting requirement (I4C/CERT-In) and formal warnings about AI-agent cyberattacks. Reversal indicator: publication of an enterprise addendum to OpenAI's data processing before the December launch.
Overall reading: the signal is mixed. Capabilities are growing faster than controls, but providers are beginning to introduce protective mechanisms; infrastructure consolidation is accelerating, but the open-source ecosystem remains competitive in price and performance.
This material was produced automatically by a large-language-model system from the public sources listed below; it is AI-generated content and may contain inaccuracies — verify facts against the original sources.
Sources
- AI-driven vulnerability discovery compresses the patch window to one day while a coordinated wave of attacks hits AI development platforms — thenewstack.io, aa.com.tr, scmp.com (+2)
- Intel shares fell sharply after an earnings report that disappointed investors on the December-quarter outlook — marketwatch.com
- Open-weight models close the capability gap and the inference market scales past 40 trillion tokens per day — venturebeat.com, asia.nikkei.com, interconnects.ai (+1)
- James Gates sues OpenAI and Sam Altman alleging ChatGPT provided specific Oleander poisoning instructions to his wife — arstechnica.com
- NVIDIA released four new open-weight Cosmos 3 Nano checkpoints for interactive world simulation and multi-sensor driving scene generation — huggingface.co
- OpenAI consolidates ChatGPT and Codex into a shared token pool, re-enables model training on Codex chats, and ships Images 2.6 — news.ycombinator.com, thenewstack.io, help.openai.com (+2)
- Saudi Aramco takes 49% stake in state-backed AI company Humain, injecting $2.6 billion in cash and AI assets — economictimes.indiatimes.com
- India's I4C and CERT-In warned banks and NBFCs of AI-agent-powered cyberattacks on payment and fund-disbursement systems — economictimes.indiatimes.com
- French regional bank Credit Mutuel Arkea said a cyberattack allowed fraudulent transactions of around 100 million euros — france24.com
- Marvell agrees to acquire Lumentum in a $115 billion all-stock deal to push optical networking for AI infrastructure — cnbc.com, marketwatch.com, finance.yahoo.com (+2)
- Intel shares fall over 28% after weaker-than-expected Q4 sales guidance citing PC market slowdown — en.yna.co.kr
- WIRED investigation reveals network of 400+ apps and domains loading hidden deepfake nudify interfaces, linked to KeLing Technology Limited — wired.com
- The BLS October jobs report showed 165,000 nonfarm payrolls added and unemployment at 4.1%, but average unemployment duration surpassed 27 weeks — cnbc.com
- MIT CSAIL researchers demonstrate GPT-6 Astra can read and edit neural network weights, drawing critical scrutiny over the claims' scope — venturebeat.com, honnibal.dev, rikiyat.github.io (+2)
- BlackX, a UEFI bootkit, infected fully patched Windows 11 machines by exploiting the Baton Drop flaw in the Windows boot manager — arstechnica.com
- Microsoft patches fail to fix on-premises SharePoint, now under zero-day attack — theregister.com (+2)
- Google released Gemini Omni 1.2 Flash with native 15-second video generation and a 30% price cut to $0.07 per second for 720p — blog.google
- OpenSSH/musl vulnerability allows passwordless SSH login via over-length password on locked accounts, fixed in OpenSSH 10.6p2 — openssh.org, arstechnica.com, sshhole.com
- Researchers assessed that Russia's Foreign Intelligence Service exploited patched iOS and Android flaws to target website visitors — arstechnica.com
- Pwn.ai published a demonstration of a VM escape from Firecracker v1.16.1 achieving root code execution on the host — pwn.ai
- Mistral released its Large 4 open-weight model 'Le Chonk' and expanded Mistral Studio with managed agents and enterprise workflows — mistral.ai, theregister.com
- DeepSeek previewed a super-agent architecture for coordinating multiple AI agents — scmp.com
- EQT buys a majority stake in Swiss cybersecurity firm Acronis at a $3.5 billion or higher valuation — theregister.com (+2)
- Mitsui Chemicals planned to acquire JSR from Japan Investment Corp. for more than $8.2 billion — asia.nikkei.com
- CXMT pushes its 4F² DRAM workaround to US export restrictions on the back of 613% revenue growth, as a Samsung leaker gets five years — scmp.com, cnbc.com