AI Weekly Malaysia

Cloudflare is merging Workers AI (its GPU-hosted inference service) and AI Gateway (its proxy for any model provider with observability, logging, and security) into a single control plane. Developers now use one unified Workers binding (`env.AI.run()`) and one REST API endpoint (`/ai/`) to route to any provider, including Cloudflare's own hosted models like `@cf/zai-org/glm-5.2`, with a built-in 'default' gateway for automatic logging.

Why: If you build AI agents or inference pipelines on Cloudflare Workers, you no longer need to choose between the Workers AI binding and AI Gateway setup—both paths are now one, and you get observability and logging by default without configuring a separate gateway. This simplifies multi-provider routing and billing consolidation, but it is a platform-specific convenience, not a reason to migrate if you are not already on Cloudflare.

Cloudflare beta-launched Radar Researcher, an AI tool that lets users query Cloudflare Radar's internet traffic datasets in plain language and receive interactive charts in response. The tool is built on Cloudflare's developer platform and grounds its LLM responses in Radar's free API data, covering DNS queries from 1.1.1.1, HTTP traffic, network quality data, and more. It is available on every Radar page and was announced as part of Cloudflare's Agents Week.

Why: Radar Researcher is a working example of the pattern where an LLM agent is grounded in a real API rather than free-form generation—useful both as a tool you can try today for internet traffic analysis and as a reference architecture if you are building AI agents over your own datasets. If you currently use Radar's API manually for network or outage research, you can now skip filter selection and API documentation and ask questions directly, though it is in beta so verify outputs against raw API calls for anything critical.

An 18-year-old use-after-free bug in Linux's SCTP networking code (CVE-2026-64564, dubbed 'SCTPhantom') allows local users to gain root and escape containers. Tencent Zhuque Lab demonstrated container escape on Debian 13, Ubuntu 24.04, Rocky Linux 9, RHEL 9, and OpenCloudOS using default seccomp without CAP_NET_ADMIN. Fixes shipped August 3 in kernels 7.1.6, 6.18.42, 6.12.101, and 6.6.148; no public exploit code exists yet and it's not in CISA's KEV catalog.

Why: If you run multi-tenant container platforms or shared Linux hosts with SCTP enabled, patch your kernels now—the bug has been present since Linux 2.6.25 (2008) and the container escape path works without special capabilities. If SCTP isn't reachable on your hosts, exposure is limited since the flaw is local-only, not remote.

Security researcher Malcolm Stagg disclosed NatJack, a new attack class presented at Black Hat USA 2026 that manipulates NAT connection-tracking state to hijack active TCP sessions, spoof DNS responses, expose mapped ports, and exhaust NAT tables. Two CVEs were assigned: CVE-2026-56181 (CVSS 8.3) in Windows NAT used by Hyper-V and CVE-2026-63913 (CVSS 8.2) in Linux Netfilter conntrack, with patches available for both.

Why: If you run multi-tenant workloads behind shared NAT (e.g., Hyper-V VMs or Linux containers on the same host), apply the Windows and Linux updates now and stop assuming hosts behind the same NAT cannot interfere with each other's connections. The attack requires privileged access on a system behind the same NAT, so isolate untrusted workloads from trusted systems sharing NAT infrastructure and encrypt internal traffic rather than relying on NAT boundaries for trust.

Arctic Wolf Labs reports a widespread AitM phishing campaign hijacking Microsoft 365 accounts via voicemail-themed emails, using a six-stage redirect chain through Google Meet, Google Ads, and Amazon S3 to bypass reputation filters. Compromised sessions are maintained at roughly eight-hour intervals using residential proxies, and attackers target employees in financial workflows to reroute payroll payments. Hundreds of organizations were hit last month across healthcare, education, manufacturing, government, and professional services in the U.S., Canada, and Europe.

Why: If your organization runs on Microsoft 365, MFA alone does not stop this attack—AitM proxies capture credentials and MFA codes in real time. Founders and IT leads should verify whether conditional access policies (device trust, impossible-travel detection, session length limits) are enforced, since the attackers maintain sessions for eight hours and use residential proxies to mimic legitimate sign-ins. The abuse of Google Meet redirect URLs and S3-hosted HTML as redirect hops means email filters that trust Google/S3 domains will not catch the initial lure.

Anthropic is hiring engineers to co-design custom ASIC inference chips, with Samsung reported as the manufacturing partner, joining Amazon, Meta, OpenAI, and Google in building in-house AI silicon to reduce dependence on costly Nvidia GPUs. The move is framed around making inference workloads more efficient and potentially profitable as chip supply remains constrained.

Why: For builders shipping on Claude APIs, this signals that inference cost-per-token could drop materially over the next 2-3 years as custom silicon replaces Nvidia GPUs in Anthropic's data centers. If you're architecting AI agents or SaaS products with tight margins on API costs, factor in that pricing leverage may shift — but don't change anything today since this is early-stage hiring, not a shipping product.

SK Hynix will invest 54 trillion Korean won ($38.1 billion) to build two new memory chip plants: a 35.2 trillion won fab in Yongin called 'Y2' and a 19.1 trillion won facility in Cheongju called 'M17.' The expansion is driven by surging memory prices caused by supply shortages and high demand from AI infrastructure builders and chip firms like Nvidia for high-bandwidth memory (HBM).

Why: Memory prices are climbing due to HBM supply tightness, which directly affects GPU availability and cloud compute costs for AI/ML workloads. Builders scaling AI inference or training should expect continued upward pressure on hardware and cloud pricing through the multi-year buildout timeline, and factor this into infrastructure budgeting and vendor lock-in decisions.

U Mobile launched its Enterprise Innovation Platform (EIP) Hub, a sandbox facility built on its 5G-Advanced infrastructure where Malaysian enterprises can develop, test, and demonstrate 5G-A and AI solutions before scaling. The hub has 15 partners including AWS, Huawei Malaysia, Palo Alto Networks, Qualcomm, and Agibot, and is working with MRANTI and Cradle Fund to connect startups with funding pathways and testing environments.

Why: If you are building AI, robotics, IoT, or edge-compute solutions in Malaysia, the EIP Hub offers a controlled 5G-A testbed and a potential route to Cradle Fund and MRANTI support—worth exploring before investing in your own infrastructure. The partner list (AWS, Qualcomm, Huawei, Agibot) signals what tech stack the sandbox is optimized for.

A weekly security roundup covering 30+ threats, including an npm supply-chain campaign ('Flooding Dropper') flooding the registry with 850 malicious packages, SideWinder's new phishing chain abusing ClickOnce files to deliver Rust-based backdoors via Cloudflare Workers, a US Congressional report on Chinese telecom presence in US infrastructure, and various RCE and one-click takeover vulnerabilities in consumer-facing products.

Why: The 850-package npm flooding campaign is the most directly actionable item: if your projects pull from npm without lockfile pinning or provenance checks, you are exposed right now. The SideWinder chain's use of Cloudflare Workers for C2 is a reminder that serverless platforms you build on can also be abused for malware hosting — relevant if you're shipping agent or automation tooling that trusts external endpoints.

AmInvestment upgraded ViTrox Corporation to 'Buy' with a RM10.25 target after 2QFY2026 revenue hit RM375M (+104.8% YoY) and net profit tripled to RM85M, driven by AI infrastructure spending flowing through the semiconductor supply chain. The research house raised FY2026-FY2028 earnings forecasts by 96%-121%, noting ViTrox's largest customer is now a hyperscaler contributing ~10% of group revenue, and that demand appears structural rather than temporary front-loading. Capacity expansion continues at Batu Kawan, Penang (Campus 4 and 5), with a renewed Pioneer tax incentive pushing effective tax rate below 10%.

Why: For Malaysian builders and founders, this signals that AI infrastructure capex is creating durable demand in the local semiconductor equipment supply chain, not just a cyclical blip. If you're building or investing in Penang-region hardware, manufacturing tech, or supply-chain SaaS, ViTrox's hyperscaler relationship and capacity expansion at Batu Kawan are concrete indicators of where AI-driven spending is landing locally.

WIRED reports that Meta approved and ran over 50 ads containing AI-generated child sexual abuse material across Facebook, Instagram, Messenger, and Threads between November 2025 and early August 2026, with some reaching several thousand accounts in the US, UK, and Europe. Researchers at the Tech Transparency Project found the ads in Meta's own ad library, many linking to 'nudify' apps, with some specifically targeting men and running for several days before removal.

Why: If you build any platform with user-generated content, ads, or AI-generated media pipelines, this is a concrete demonstration that automated content review at Meta's scale still fails catastrophically on the most obviously prohibited content. The ads were not buried user posts—they passed Meta's ad review and collected ad revenue. Anyone shipping AI image generation features or ad marketplaces should treat trust-and-safety moderation as a hard dependency, not an afterthought, and should assume their automated filters will miss abuse that seems glaringly obvious to a human reviewer.

Cloudflare launched a Billable Usage API for self-serve accounts, exposing a single endpoint that returns usage and cost data broken down by product (Workers, R2, D1, Workers AI, Vectorize, Images, Stream) and charge period. Response fields map directly to the FinOps Open Cost and Usage Specification (FOCUS) columns, so teams already ingesting FOCUS data from other providers can integrate with familiar semantics. Data updates daily, with real-time updates planned.

Why: If you run AI agents or automation that provisions Cloudflare resources (Workers, R2, D1, Workers AI), you can now poll one endpoint to catch runaway spend before the monthly bill arrives instead of relying on the dashboard. Teams already using FOCUS-compatible FinOps tooling can pipe this into existing cost pipelines without custom mapping.

Apple CEO Tim Cook is stepping down, and App Store head Phil Schiller has exited his role, staying on as an Apple Fellow. Day-to-day App Store operations will be handled by Carson Oliver, a 14-year App Store veteran, and the App Store is being moved from the marketing department to the services division under Eddy Cue. This follows other high-level departures including COO Jeff Williams and VP Lisa Jackson.

Why: If you ship iOS apps, the App Store moving from marketing to services under Eddy Cue could eventually shift review priorities, fee structures, or policy enforcement — but no concrete changes are announced yet. The ongoing Epic legal context (external payment link compliance) means any new leadership could reinterpret how strictly those rules are applied. Monitor for policy changes rather than acting now.

The Malaysian government's 'AI Untuk Rakyat' programme offers Malaysians aged 18-30 three months of ILMUchat Pro (worth RM150) for free after completing AI training on the Rakyat Digital platform. The programme targets 100,000 young Malaysians and aligns with PM Anwar Ibrahim's recent announcement on AI adoption.

Why: If you're building AI products or SaaS in Malaysia, this signals the government is actively subsidising distribution of a local AI chat tool to 100,000 young users — a potential competitor or complement to your own product. Founders should note the Rakyat Digital platform as a distribution channel and consider whether government-backed AI literacy programmes create adjacent opportunities. For those in the 18-30 demographic, it's a free local AI tool to evaluate against ChatGPT or Claude.

FreeCORE is an independently maintained continuation of TrueNAS CORE, built on FreeBSD and OpenZFS, after the original TrueNAS CORE lineage ended at 13.3. The current stable release is 15.0-U1, and existing TrueNAS CORE 13.3 systems can upgrade in place via a simple enrollment script. Source code is hosted on Codeberg with a GitHub mirror.

Why: If you run TrueNAS CORE 13.3 in production or homelabs, you now have a migration path beyond end-of-life via in-place upgrade to FreeCORE 15.0 — evaluate whether to adopt this community fork or move to TrueNAS SCALE/another platform before 13.3 stops receiving security patches.

GrapheneOS reports that the Pixel 11 will no longer support ARM's hardware Memory Tagging Extension (MTE), a feature present on Pixel 8 and 9 that catches memory safety bugs at the hardware level. The HN discussion drew 316 points and 189 comments, indicating significant developer concern about the regression.

Why: If you build or test Android apps on Pixel hardware and relied on MTE for catching heap memory corruption during development or fuzzing, Pixel 11 is no longer a viable target for that workflow. Security-focused teams should factor this into device procurement decisions and may need to keep older Pixel units for MTE-based testing.

Samsung presented their LPDDR5X-PIM architecture at Hot Chips 2026, placing MAC units inside each of 16 DRAM banks on standard LPDDR5X-9600 chips to exploit internal bandwidth of 614 GB/s versus 76.8 GB/s for normal external access. Each PIM block contains a MAC tree with 1024-bit instruction registers and supports INT8, FP8, and 4-bit formats, yielding 2.4 TOPS per chip at 4-bit; eight chips together reach 9.6 INT8 TOPS, roughly matching Intel Meteor Lake's NPU, but requiring 128 GB of LPDDR5X. Crucially, PIM operations are invoked using standard LPDDR5X protocol commands, meaning no exotic memory controller is needed.

Why: For builders working on edge AI or mobile inference, PIM could eventually reduce the memory-bandwidth bottleneck that limits large-model deployment on constrained devices—but at 2.4 TOPS per chip and with no shipping product timeline given, there is nothing to adopt or design around today. The most actionable detail is that Samsung is keeping the interface standard LPDDR5X, which means future PIM-aware software stacks could target existing memory controllers rather than requiring custom hardware.

Micron workers are increasingly supporting a strike over bonus pay, with a labor union pushing for a profit-sharing scheme. The discontent is fueled by comparisons to Samsung and SK hynix employees who receive bonuses worth hundreds of thousands of dollars.

Why: Micron operates major fabrication and assembly facilities in Malaysia, so labor disruption at Micron could affect memory chip supply lines that Malaysian builders and hardware-dependent startups rely on. If a strike materializes, expect potential DRAM/NAND price volatility and delivery delays.

PT Infra Fiber Teknologi (IFT), a joint venture between Arsari Group and Indosat Ooredoo Hutchison, has launched RAIA Grid—an 86,000-km open-access fibre network spanning Indonesia that connects 5G, FTTH, and data centre-to-data centre links. The platform claims to embed machine learning and agentic AI into network planning, deployment, and operations, targeting telcos, cloud providers, hyperscalers, and enterprises. The launch is positioned as foundational infrastructure for Indonesia's AI and compute growth.

Why: For Malaysian founders or developers expanding into Indonesia, RAIA Grid's open-access DC-to-DC fibre model could reduce barriers to multi-region cloud and data centre deployments there—worth tracking if you're evaluating Indonesian infrastructure partners. However, the AI-powered operations claims are vague with no technical detail, so treat the 'agentic AI' framing as marketing until specifics emerge.

A weekly cybersecurity roundup covering a 296K-device IoT botnet, 100+ water systems targeted, and a SharePoint RCE chain. The most detailed incident is a failed extortion attack on ReliaQuest on August 22, 2026, where attackers registered a lookalike domain, stood up a fake SSO page behind a CDN, called employees impersonating security staff by name, and got one employee to enter their password and approve an MFA push notification—granting brief view-only session access before being contained.

Why: The ReliaQuest incident lays out a concrete, repeatable attack pattern your SaaS or internal tools should defend against: lookalike domains registered and burned within the hour, fake SSO pages behind CDNs, and MFA push fatigue abuse. If you ship SSO or MFA, consider number-matching MFA, conditional access policies that block unfamiliar IPs/CDNs, and domain monitoring for lookalike registrations. The SharePoint RCE chain is relevant only if you run SharePoint infrastructure.

HP used its Q3 earnings call to pitch AI PCs as a cost-saving alternative to cloud AI tokens, with Personal Systems president Ketan Patel citing token economics, data governance, and agent management as drivers. AI PCs made up 46% of HP's PC mix this quarter, projected to hit 60-70% by fiscal 2027 and over 70% by 2028, with HP openly acknowledging the shift is 'margin accretive.'

Why: If you're paying per-token for cloud inference on workloads that could run locally, the cost math may now favor edge deployment—especially for repetitive internal tasks like customer service or productivity assistants. Evaluate whether your current cloud API spend on stable, smaller-model workloads justifies the capex of AI-capable hardware instead.

Anthropic is previewing the Model Hardware Standard (MHS), a model-agnostic specification using Model Context Protocol (MCP) that allows AI agents to operate physical lab and manufacturing devices like microscopes and robotic arms. It aims to reduce hardware integration time from months to hours by providing a standardized driver with simple read/write primitives. The research preview is currently limited to scientific and advanced manufacturing partners, with plans to make it open source later.

Why: If you build AI agents using MCP, this signals an expansion of the protocol into physical hardware control, though the current preview is restricted to scientific and manufacturing partners. General builders should monitor how MCP standardizes physical device interaction but do not need to change current software stacks yet.

Qwen released Qwen3.8-Flash-Next, a multimodal Mixture-of-Experts model with 125B total parameters but only 6B active, described as an early preview of the Qwen4 architecture. Simon Willison tested quantized versions (72.5GB UD-IQ1_S and 78.9GB UD-Q2_K_XL) on a DGX Spark via Unsloth, sharing early image generation results.

Why: If you're running local models, the 6B active parameter count means this could run reasonably on high-end consumer hardware despite the 125B total size—but the quantized weights are still 72-79GB, so you need serious VRAM. The MoE architecture preview for Qwen4 is the signal worth watching if you're choosing model families for agent pipelines.

Flipboard has acquired Graze, a Portland-based startup that lets creators build, customize, and monetize Bluesky feeds using contextual (non-tracking) ads with a 70/30 revenue split favoring the feed creator. Graze has sent over 41 billion posts to ~12 million people across 7,000+ feeds since launching 21 months ago, with ~60% of traffic currently monetized. Flipboard CEO Mike McCue highlighted the privacy-preserving ad model and revenue-sharing economics as the key attraction.

Why: If you're building on open social protocols like Bluesky's AT Protocol, Graze's contextual ad model offers a concrete monetization path that doesn't rely on user tracking — worth evaluating if you're curating feeds or building apps on AT Protocol. The 70/30 split and 60% monetization rate are real numbers to benchmark against if you're considering feed-based revenue models.

The article argues that traditional SOCs are bottlenecked by human-in-the-loop alert queues and proposes an inverted model where AI agents investigate signals first—validating detections, profiling entities, correlating activity—then escalate evidence-backed verdicts to humans who judge output rather than conduct investigations. Agents run asynchronously and in parallel, turning threat hunting from a capacity-limited human task into a continuously scalable process.

Why: For anyone building AI agent workflows, the 'investigate-then-escalate' inversion is a reusable design pattern beyond security: let agents do parallel, structured triage against a playbook and surface only evidence-backed decisions to humans. If you ship agentic products, consider whether your architecture lets agents run multiple async investigations and return verdicts rather than queuing everything for human review.