Monday, November 2, 2026

Sophos found a ransomware crew handing VPN logins to Cursor agents, Cisco Talos found an infostealer that falls back from Claude to Kimi to Gemini when one model refuses, and Anthropic is tightening what Claude Code on Opus 5.5 will do. Meanwhile the classic exposures keep paying out: over 1,000 GitLab servers remain unpatched and under attack.

TLDR

Over 1,000 GitLab servers are still unpatched against CVE-2026-97229 (Nov 2), a CVSSCommon Vulnerability Scoring SystemA standard 0 to 10 severity score: 0.1–3.9 low, 4.0–6.9 medium, 7.0–8.9 high, 9.0–10 critical.FIRST ↗ 10.0 remote code executionRemote code executionA flaw that lets an attacker run malicious code on a target's computers or network.Cloudflare ↗ flaw that CISACybersecurity and Infrastructure Security AgencyThe U.S. cyber defense agency, which works to reduce risk to cyber and physical infrastructure. It runs KEV.CISA ↗ lists as exploited. Security Infra
Langflow 1.0.0 through 1.12.8 has two unauthenticated critical flaws (Nov 2); the only fix is 1.13.0. Security Agents
A ransomware crew drove intrusions through Cursor agents, and a new infostealer routes victim data through Claude (Nov 2); both use AI as the operator. Security Agents
Claude Code on Opus 5.5 now refuses to clone specific games and some reverse engineering (Nov 2); Anthropic confirmed the game rule. Agents Policy
Marvell agrees to buy Lumentum in a $115 billion all-stock deal (Nov 2), one of the largest chip mergers on record. Business Infra

Upcoming Deadlines

Date	What	Who it hits	Do this	Source
Wed Nov 4	Moonshot promises Kimi K3.1 Pro full open weightsA model whose trained weights are published for anyone to download and run.NTIA ↗ (The Decoder, Oct 23)	Self-hosters of open models	Line up an eval; read the license	The Decoder
Wed Nov 4	Apple Intelligence beta with Alibaba Qwen starts in mainland China (iOS 27.3 wide release in December)	iOS developers and teams shipping in China	Test Siri, Writing Tools and Visual Intelligence flows against the Qwen-backed beta	TechNode
Mon Nov 9	SEC roundtable on machine-understandable data and "Super Intelligence"	Public-company and data teams	Watch for stream	SEC
Tue Nov 10	CISA KEVKnown Exploited VulnerabilitiesCISA's list of vulnerabilities with reliable evidence of active exploitation and a clear fix, used to set patching priority.CISA ↗ federal patch deadline for Brotli CVE-2026-111497 (Node2Shell)	Anyone running Node.js, .NET, Bun, Envoy or other servers with a vulnerable Brotli decoder	Patch Brotli-dependent runtimes and proxies; confirm transitive copies	BleepingComputer
Thu Nov 12	Rust 1.100 stabilizes Cargo minimum publish age (dependency cooldown, off by default; Socket, Oct 20)	Rust teams	Set a cooldown in Cargo config	Socket
~Thu Nov 12	OpenAI's proposed end of model access in Cursor (official date not confirmed; Cursor may end it sooner)	Cursor users on OpenAI models	Check Cursor terms	OpenAI
Mon Nov 16	GitHub Copilot retires Claude Opus 4.8, Claude Opus 5, GPT-5.6 Terra (replacements Opus 5.5, GPT-6.1 Sol / GPT-6 Luna)	Copilot users with pinned models	Repoint model pins	GitHub

Action Items

Check every self-managed GitLab instance today for version 19.4.2, 19.3.5 or 19.2.8 or later. If one sat internet-exposed and unpatched after Oct 12, follow Alpine Linux and rebuild it from scratch rather than trusting missing logs.
Upgrade every Langflow server to 1.13.0 and confirm LANGFLOW_AUTO_LOGIN is off, or take the server off the internet. IBM offers no workaround.
Put spend caps on your Google Cloud sandbox and agent-playground projects this week using the new spend caps API, and remember Provisioned Throughput is never halted by a cap.
Run your reverse-engineering and game-development Claude Code workflows on Opus 5.5 this week, and keep a second model or harness configured if refusals break them.
Strip white-on-white and zero-size text from third-party documents before they reach any LLMLarge language modelA neural network trained on huge amounts of text that generates text, answers questions and summarizes.Anthropic ↗ review pipeline.

Security

GitLab CVE-2026-97229: over 1,000 servers still unpatched while attacks continue (Nov 2) Security Infra Critical · CVSS 10.0 · Exploited in the wild (CISA KEV, federal deadline Nov 1) · Patched in 19.4.2 / 19.3.5 / 19.2.8 Update — Shadowserver counts more than 1,000 exposed servers still on vulnerable builds, three weeks after the Oct 12 fix. Alpine Linux says attackers probed its GitLab on Oct 20-21, crashing processes repeatedly. It had installed 19.3.4, not a patched build, and with missing logs it rebuilt the whole instance. (BleepingComputer | BleepingComputer)
Langflow CVE-2026-93680 and CVE-2026-110186 allow unauthenticated bypass and code execution (Nov 2) Security Agents Critical · CVSS 9.8 (IBM) · Exploitation unknown (not in CISA KEV as of the Oct 30 update; a public Metasploit module covers the auth bypass) · Patched in 1.13.0 IBM's bulletin lists every release from 1.0.0 through 1.12.8, and 1.12.8 shipped the same day as the fix. One path reaches code execution through MCPModel Context ProtocolAn open standard that lets AI apps connect to data sources, tools and workflows through one common interface.MCP ↗ tool enumeration. Version 1.13 turns automatic login off by default; the module author lists a narrower affected range than IBM, so patch regardless. (The Hacker News)
Ransomware crew and a Kaspersky-tracked campaign hand intrusion work to AI agents (Nov 2) Security Agents High · CVSS: not yet assigned (no CVECommon Vulnerabilities and ExposuresA public catalog that gives each known security vulnerability a unique ID, a short description and at least one reference.NIST ↗; scanners will be silent) · Exploited in the wild · No patch; restrict and monitor agent tooling Sophos recovered Cursor artifacts showing the Aur0ra operators giving agents VPN logins and targets; the agents mapped Active Directory and reached domain admin in some networks, with humans keeping the break-in and ransomware step. They posed as red teamers to dodge guardrails. Kaspersky separately saw the open-source CyberStrikeAI framework take education-sector victims from recon to full domain compromise. (Help Net Security | Securelist)
VOLTA infostealer ranks victims by value using Claude, Kimi and Gemini (Nov 2) Security Medium · CVSS: not yet assigned (malware, no CVE) · Exploitation unknown · No patch; block and monitor Cisco Talos says VOLTA bundles a victim's stolen browser, Slack, cloud-key and receipt data and sends it through OpenRouter, defaulting to Claude Opus 4.6. It falls back to Kimi K2, then Gemini 2.5 Flash, when a response contains refusal phrases such as "I can't". Unexpected egress to LLM gateways from workstations is a detection signal. (The Hacker News)

Models & Tools

Claude Code on Opus 5.5 gets a game-cloning rule and users report reverse-engineering refusals (Nov 2) Agents Policy An Anthropic employee confirmed two added sentences in Claude Code's Opus 5.5 prompt steering it away from recreating a specific existing game. Users also report blocked reverse-engineering tasks starting Oct 31; one Tom's Hardware alumnus kept working by switching to Opus 4.8. The reverse-engineering rule is unconfirmed by Anthropic. Teams doing security or compatibility work should test now. (Tom's Hardware)
Meituan releases LongCat-2.5, a 1.6T-parameter multimodal coding model (Nov 2) Agents Infra About 48B parameters are active per tokenA small unit of data, such as part of a word, that AI models read and generate. Many AI services price by tokens in and out.NVIDIA ↗. Meituan says it targets long-horizon agenticDescribes AI systems where a model uses tools and decides its own steps instead of following a fixed script.Anthropic ↗ coding with native vision, and that W4A4 (4-bit weights and activations) inferenceRunning a trained model on new input to get an output, as opposed to training it.NVIDIA ↗ runs on domestic Chinese chips. The preview has been on Meituan's API since Sep 25. No benchmark figures were in the report, and weight availability is not stated. (Pandaily | TechNode)
Apple puts Alibaba's Qwen inside Siri and Writing Tools for China, beta from Nov 4 (Oct 31) Business Customer AI Apple announced the Apple Intelligence rollout in mainland China will run on Qwen for Siri, Writing Tools and Visual Intelligence. It arrives as an iOS developer beta Nov 4 and ships widely with iOS 27.3 in December. Apps that call system AI features will hit a different model in China than elsewhere. (TechNode)

APIs & Pricing

Google Cloud spend caps reach general availability with an API and automatic monthly resets (Oct 30) Pricing Infra Caps for AI services trigger within minutes on estimated live costs, and the spend caps API sits in the Cloud Billing Budgets API so a new sandbox can get a cap on creation. Pay-as-you-go and committed-use requests halt at the limit; Provisioned Throughput never does. Default alerts fire at 50%, 80% and 100%. (Google Cloud)

Industry & Business

Marvell to buy Lumentum for $115 billion in stock (Nov 2) Business Infra The Information reports Marvell, which designs data center, networking and storage chips, agreed Monday to acquire the laser maker in an all-stock deal valuing Lumentum at $115 billion, one of the largest semiconductor mergers ever. (The Information)
Saudi Aramco buys 49% of Humain for about $2.6 billion (Nov 2) Business Infra Aramco contributes roughly $2.6 billion in 2026 plus its own AI technology to PIF-founded Humain, which plans three gigawatts of data center capacity by 2030. Humain says the Iran war has not cut demand and is hardening its data centers. (The Decoder)
OpenAI signs Abu Dhabi deals and courts Samsung and SK as its $20 billion round leans on Gulf money (Nov 2) Business Infra ADNOC and G42 will deploy OpenAI technology, and Mubadala signed an investment-analysis agreement, per Abu Dhabi's media office. Altman also met about 50 Samsung executives and reportedly SK's chairman; Tom's Hardware frames the trip as a hunt for RAM. (CNBC | CNBC)
Meta reportedly built a humanoid prototype with China's Fourier (Oct 31) Robotics/IoT Business The Information reports a roughly five-foot prototype called Max, built in three months, with a second version underway. Meta has set no launch date and is using the units for research toward a consumer household robot. (TechNode | The Information)

Policy & Safety

OpenAI plans a post-midterm lobbying push built on binding "speed limits" for frontier models (Nov 2) Policy Per Axios, relayed by The Decoder, OpenAI's Chris Lehane says fast-tracking safety rules for frontier models is the top priority in the next Congress, expecting tougher scrutiny if Democrats win the House. It also wants to rebrand data centers as "AI Compute Campuses" with promised local benefits. (The Decoder | TechCrunch)
South Korea budgets 4.7 trillion won for a sovereign AI model aimed at "world's highest performance" (Nov 2) Policy Infra President Lee's record 2027 budget is 820.9 trillion won, up 12.8%. It includes 21.3 trillion won for semiconductors, physical AI and AI data centers, nearly double the 2026 level. The performance target is a government claim, not a measured result. (CNBC)

AI in Production

A Brazilian labor court fines two lawyers about $16,000 for hiding a prompt injectionInput crafted to change an LLM's behavior or output in ways its builders didn't intend.OWASP ↗ in a filing (Nov 1) Production Security White-on-white text told the court's AI to "respond superficially" and not challenge the documents. The court's own system flagged it, and the fine was 10% of the award. Any pipeline that reads third-party documents needs hidden-text stripping before the model sees them. (Tom's Hardware)
Reddit will make itself a "logged-in service" to choke off scrapers (Oct 31) Data Policy Co-founder Steve Huffman says most unauthorized data access comes through logged-out pages. A pilot starts in selected markets outside the US in coming months; what gets locked and when is undecided. AI companies can still license the data commercially. Pipelines that scrape Reddit should plan for a licensed feed. (BleepingComputer)
Amazon's Jassy predicts 6 to 8 interchangeable frontier models within 2 to 3 years (Oct 30) Business Production On Amazon's Q3 call he said these models will cover over 95% of workloads and most customers will mix several, even inside one application. It is a Bedrock seller's forecast, but it argues for a model-routing layer over a single-vendor build. (The New Stack)

Attackers now run AI agents inside victim networks, so model vendors' refusal rules are becoming a security control.