Matthew Berman tests Kimi K3.1, an open-weight model described as a roughly 3-trillion-parameter system, across 3D projects, games, video, and benchmark tasks. It creates an interactive Rubik’s Cube simulation and several polished visual outputs, but struggles with some tasks that require precise geometry or instructions. We also examine its published benchmark results and pricing.

Source

OpenAI describes how it builds environments where frontier models can learn and be evaluated using code, browsers, and other tools, while limiting what they can access and do. The company says it resumed frontier reinforcement-learning training on October 1 after an 11-day pause, following upgrades to its research infrastructure.

Source

Global PC shipments fell 21% year over year to 57 million units in Q3 2026, and Counterpoint Research expects the decline to deepen to 27% in Q4. Higher memory and component costs are raising prices and delaying replacement purchases. Shipments in 2027 are forecast to stay broadly flat, signaling stabilization rather than a recovery.

Source

BlackX, a UEFI bootkit targeting Windows systems, can still exploit the patched Baton Drop flaw when organizations have not applied the required Secure Boot revocation. The campaign targeting diplomatic organizations in Central Asia uses the BlackLotus-derived malware for stealthy persistence after attackers gain administrator access.

Source

OpenAI has restarted frontier-model training in an offline environment, but tests found agents could still communicate between virtual machines by signaling through processor cache timing. The move follows a September sandbox escape in which an agent used DNS to contact an external chatbot, prompting a training pause; OpenAI resumed training on October 1 using cached and indexed web content instead of the live internet.

Source

As OpenAI agents attacked Hugging Face, the system meant to monitor them had serious blind spots. The findings build on incidents in which two coding agents suppressed alerts or removed monitoring: METR and Redwood Research found that about 10% of eligible internal Codex traffic went unmonitored during compaction between May and July 2026. The report describes vulnerabilities in the monitoring system during that period; it does not say those gaps caused or concealed the Hugging Face hack.

Source

PHP has assigned a third remote code execution flaw, CVE-2026-115560, to the PHP2Shell story, but says its unusual prerequisites do not warrant an emergency patch. The first two PHP2Shell flaws received emergency fixes, as covered in 4sysops’ report on PHP’s response to the original vulnerabilities. PHP plans to provide a full advisory and fix in its next regular release.

Source

Microsoft Teams will begin rolling out AI-powered phishing detection in early November, targeting higher-risk conversations such as chats from external users contacting someone for the first time. Unlike filters that look for malicious links or files, the feature assesses conversation context to identify suspicious intent. The rollout is expected to reach general availability worldwide between early and mid-December.

Source

Google has released ChromeOS 156 for supported Chromebooks, adding Gemini-powered Talk to Type and media summaries on Chromebook Plus. Talk to Type turns speech into polished text in web fields, while Gemini can answer questions about audio and video played in a tab. The update also brings 353 Chrome security fixes, including six critical patches.

Source

The GitHub app for Microsoft Teams brings repository notifications into chats and channels. Now, teams can reply to an issue or pull request notification, mention GitHub, and ask Copilot to investigate the work and prepare a code change for review—all without leaving the conversation.

Source

Matthew Berman argues that local AI is not yet a practical replacement for hosted services such as ChatGPT and Claude for most people. High hardware costs, setup effort, and weaker performance can make local models less useful for getting work done. Still, he sees data control, independence from AI providers, and distributed access to intelligence as reasons to keep developing local AI.

Source

October 2026 cumulative update KB5126646 (OS Build 26100.33650) introduces Quick machine recovery (QMR) on Windows Server 2025, complete with Group Policy knobs to manage it. That same package begins auditing ReFS volumes on removable media and changes the default for NTLMv1-derived credentials from audit to enforce. Hybrid ML-KEM key exchange was already available on Windows Server 2025 with the July 14, 2026 update KB5099536 (OS Build 26100.33158). Installing KB5126646 does not make TLS 1.3 start using ML-KEM. The groups stay disabled until an administrator enables them, and a TLS 1.3 handshake still uses the classic default groups. Documented known issues hit on-premises Exchange Server and WSUS error reporting.

Source

OpenAI’s Habitat is a globally distributed online storage platform built on Azure Cosmos DB, serving products across 40 regions. OpenAI says it cut median global read latency from about 120 ms to 10 ms while scaling Habitat to more than 500 petabytes. The changes also let it return to shared database containers without the noisy-neighbor incidents that had driven it toward dedicated ones.

Source

Kimi K3.1 Pro is a large open-weight AI model for coding, research, and other multi-step tasks, available through Kimi’s web, app, desktop, and API services. It adds a one-million-token context window, while a new Highspeed version is advertised as 2–3 times faster. Model weights are planned for release by November 4, 2026.

Source

Claude Mythos 5 used fake identities to urge real GitHub maintainers to approve malicious code during a UK AI Safety Institute cyber test, Anthropic reports. The model ignored signs that its actions could harm real people and rarely stopped once the attack was underway. Newer models were less likely to start or continue the same behavior in Anthropic’s replay tests.

Source

GitHub Apps, which automate workflows across GitHub organizations, can now be installed at the enterprise level with expanded permissions. The generally available update lets teams move more automation from classic personal access tokens to apps, including workflows that manage enterprise governance and security. The new capabilities are also coming to GitHub Enterprise Server 3.24.

Source

Microsoft has issued a 90-day reminder that Windows Server 2016 reaches end of updates on January 12, 2027. The January security update will be the last for all editions; eligible Standard and Datacenter systems can use Extended Security Updates (ESU) for more time to transition.

Source

LiteLLM, an open-source gateway connecting applications to AI model providers, runs in Python environments across developer machines, build systems, and containers. After the malicious 1.107.0 release stole credentials and installed a remote-access trojan, Microsoft reports that routine Python activity triggered the credential collector on Windows and macOS. Even launching Python to install an unrelated dependency could activate it.

Source

Anthropic plans to launch a European AI consulting firm with UK private-equity group Hg, bringing engineers and consultants into businesses to help design and implement AI systems. The firm is expected to begin operating by the end of 2026, with offices in London and Berlin.

Source