AI Weekly Malaysia

Anthropic is giving six METR researchers employee-level access—company laptops, office space, and internal tools roughly on par with Anthropic’s own risk-assessment teams—to independently assess its models, safety practices, and risk management. METR can publish its own assessments and Anthropic says it cannot veto publication, but Anthropic can require redactions for customer data, attorney-client privileged material, and model-architecture information, and METR may publicly characterize those redactions. The arrangement comes amid disclosures every week or two since late August of AI agents acting outside their instructions, and METR is the first evaluator at the top of a three-tier system Anthropic is building.

Why: If you build on or standardize around Anthropic models, METR’s future reports could become a vendor-risk input, but the redaction carve-outs for model architectures and customer data mean this is not full architectural transparency. Ask what METR will actually be able to publish before treating this as independent assurance.

CNBC reports that demand for "forward deployed engineers" (FDEs) — engineers who embed with enterprise customers to make AI systems produce measurable business results — is surging: U.S. FDE job postings topped 7,300 in the first nine months of 2026, up from fewer than 200 in 2023, per labor analytics firm Lightcast. AWS has committed $1 billion to its forward deployed engineering initiative, and Anthropic plans to train 10,000 FDEs by the end of 2027. The article frames this as Big Tech's answer to investor skepticism that hyperscaler AI spending outpaces any foreseeable payoff, following a rough few days for AI-themed stocks.

Why: The hiring signal here points at the implementation layer, not model-building: 7,300+ U.S. FDE postings in nine months versus fewer than 200 in all of 2023 means the paid work is increasingly 'make this LLM work inside a specific company's process.' If you're choosing what to learn or what your studio/agency sells, enterprise integration plus AI depth is where the advertised demand is, and AWS's $1B and Anthropic's 10,000-FDE target suggest partners and tooling will be pulled along. Note the article gives only U.S. posting data — it offers no Malaysia or Southeast Asia figures, so treat any local opportunity claim as unverified by this text.

Cloudflare has open-sourced Void Platform, Void SDK, and Void CLI as a single JavaScript package and Vite plugin installable with `npm install void`. Void adds server routing, typed data access, and deployment to Cloudflare Workers for Platforms, with admin controls for access, resource limits, and deployment/usage inspection. The post says this can cut internal app shipping from days/weeks to minutes and support non-technical teammates using coding agents, after VoidZero joined Cloudflare and the plan shifted from a private platform to open source.

Why: If your team already uses Vite and Cloudflare Workers, evaluate `npm install void` as a way to let vibe-coded/internal apps ship through a shared platform with resource limits and access controls instead of giving every employee broad cloud access; teams not on that stack can treat it as a forkable reference rather than an urgent migration. No Malaysia-specific policy or funding detail is present in the text, so the impact is stack-based.

Citizen Lab at the University of Toronto says Israeli ad agency Discover-Tech has served millions of ads since 2024 through its Controlium platform, which can mass-profile and deanonymize users then silently install zero-click spyware. Dozens of Al Jazeera devices were reportedly compromised, with ads shown in Yemen, Brazil, Mexico, and Czechia; separate research by Confiant says the ads also reached users on networks of European and US government entities, the UN, NATO, the ICC, and the IAEA.

Why: If you run an ad-supported app or embed third-party ad SDKs, this shows that ad delivery can be repurposed as a mass-profiling and zero-click infection channel—Controlium allegedly harvests IP and identity from a benign ad, then hands targets to spyware within seconds. Treat ad tech as untrusted third-party code and weigh whether the revenue justifies exposing your users to that supply chain.

The Hacker News reports that a critical authenticated remote code execution flaw in Langflow — the open-source low-code builder for LLM apps and agents — was targeted in 1,423 exploit attempts recorded across honeypots. The excerpt provided is truncated to the headline and site boilerplate, so it does not include the CVE ID, affected version range, patch release, or the researcher/vendor attribution behind the honeypot count.

Why: If you self-host Langflow — a common setup for teams prototyping agents on a VPS or cloud instance — this is a prompt to check your version and patch rather than wait for a full write-up, because 1,423 honeypot hits means scanning and exploitation attempts are already happening in the wild. Note that the flaw is described as authenticated, so the practical question for your own deployment is who can reach the login and whether default or shared credentials are in use; the excerpt does not give a version or fix, so verify against Langflow's own release notes before assuming you are covered.

Meta released Muse Spark Open, an open-weight multimodal mixture-of-experts model on Hugging Face with roughly 816B total parameters, ~42B active per token, and a context window up to 1M tokens. It accepts text, images, audio, and video as input but outputs text only, and Meta says it was trained for agentic coding tasks like multi-file refactors, long debugging sessions, tool calling, and browsing with images. Checkpoints ship in BF16 and NVFP4 formats under Muse-Spark-Open-816B-A42B, with 256 experts per MoE layer and 8 selected per token; it shares a tokenizer and chat template with the smaller dense Muse Glimmer released in August.

Why: The deployment-relevant detail is the NVFP4 checkpoint alongside BF16: at 42B active parameters per token and a 1M-token context, the quantized checkpoint is likely what decides whether self-hosting is feasible versus staying on an API. If your agent loop benefits from screenshot, voice-note, or screen-recording input, this is one model that handles all three in a single interface, but output is text-only, so any image or audio generation still needs separate tooling. Note that 'most capable US open-weights model to date' is Meta's own claim in this post and is not benchmarked here, and the post had only 9 upvotes at capture, so independent evaluation is still thin.

At the AI Engineer New York conference (published Oct 14, 2026), OpenAI's Joseph Kim — Head of ChatGPT for Financial Services — introduced a "threshold of usefulness," defined as the point where a model becomes capable and reliable enough to save you time, and predicted that by 2027 finance will lean less on Excel and PowerPoint while agents become good enough for "investment idea generation and trading." John Feminella, SVP Architecture at the hedge fund Two Sigma, gave a keynote on the product-design side of the same problem, and the conference's stated main lesson was that safe autonomy depends on where you put controls and human judgment, not on how many restrictions you impose. Kim framed human-in-the-loop placement as an interface question: what the user still needs to control, review, or approve.

Why: If you are building any agent that touches money, numbers, or other consequential decisions, this gives you two concrete design constraints to implement rather than debate: per-value provenance (Kim's point that in finance you must check where a number came from) and an explicit decision about which steps the user approves versus which run automatically. The 2027 prediction that agents will handle investment idea generation and trading is a signal about where competing products may aim next, not a roadmap you can rely on.

Simon Willison created "Tree Appreciator," a Three.js spot-the-difference game about two algorithmically generated trees, using the default Claude iPhone app in chat mode with Claude Opus 5.5 Max. The prompt included instructions to "Think hard" and "make sure it is both fun and beautiful"; Claude took 3 hours and 7 minutes to return a game with music, animation, and achievements. Willison notes that telling Opus 5.5 "make sure it is fun" worked "uncomfortably well," and that this was chat mode, not Claude Code mode.

Why: If you vibe-code generative UI or games, this is a concrete datapoint: a single chat-mode Claude Opus 5.5 Max prompt ran for 3h7m and produced a polished Three.js game. Try the explicit "make sure it is fun" clause on your next generative game/UI prompt, and decide whether you still need Claude Code for this kind of task or whether long chat-mode runs are enough.

A post on sunshowers.io argues that Anthropic's newly announced version control system, Scion, closely mirrors Meta's Mononoke: both use a 'derived data' concept to make operations like blames more efficient, and the author says they described Mononoke's derived data in 2022 Oxide candidate materials for a system their team built at Meta in 2017-18. The author notes Anthropic credited Sapling only for 'hard-won lessons' and did not mention Mononoke at all, calling the omission close to plagiarism, while stating they have no problem with the design being copied since Mononoke is open source. The post adds that because Mononoke is open source and in LLM training sets, asking current models about prior art immediately surfaces the Mononoke similarities.

Why: If you are using an LLM to draft a system design, the model will likely reproduce prior art from its training set without naming it — the author demonstrates this by getting current models to point straight at Mononoke. The practical move is to explicitly ask the model for prior art before you ship an architecture doc or a 'novel' design claim, because the source of the idea is now one prompt away and readers like this one will find it. It also means teams evaluating Scion as a VCS can read Mononoke's open-source design for the derived-data approach rather than waiting on vendor documentation.

Cloudflare analyzed HTTP traffic to two Yandex ASNs after reported Ukrainian drone strikes hit three Yandex facilities between October 7 and 11, 2026. Yandex Search (AS13238) traffic fell to 99% below baseline on October 11 and Yandex.Cloud (AS200350) to 90%, with both still 86-90% below normal through October 13 and Yandex overall at 85% of normal as of October 14 18:00 UTC. Crucially, the Search ASN drop was almost entirely verified crawlers (93.6% of its bot requests were Verified, 99.5% of all requests automated), while Yandex.Cloud saw a decline in vulnerability scanning and probing but high-volume automation migrated to other hosting providers rather than disappearing.

Why: If you block or rate-limit by ASN, IP range, or hosting provider, this is evidence that the block is fragile: when one provider goes down, the scanning and probing traffic moved elsewhere instead of stopping. The pre-event baseline — automated requests at 36.1% of global HTTP traffic and 35.8% from Russian IP addresses — also means aggregate traffic charts are a bad proxy for whether abuse stopped. Build your abuse detection on request behavior (the classification approach Cloudflare used), not on which network the traffic came from, and don't treat an ASN's traffic drop as attribution of who is behind it.

Android Authority reports that Google now automatically verifies and attaches your phone number to every Google account signed in on an Android device, with no explicit confirmation prompt. The control, 'Verify phone numbers automatically,' sits in Android's Google settings and is opt-out; Google updated its documentation between August and October 2025 to describe the behavior. Switching it off does not undo anything — numbers already attached must be removed manually from each account's Personal Info page.

Why: If you keep a work Workspace, Google Cloud, or Play Console account signed in alongside your personal one on the same Android phone, that account probably now carries your personal phone number — a link that affects account recovery, identity separation, and anything your employer or a client can see on a managed account. Concretely: open Android Settings > Google, turn off 'Verify phone numbers automatically', then go account by account into Personal Info and delete any number already listed. Anyone running alt or throwaway accounts for side projects should assume those are now tied to the same number.

Simon Willison link-posts to indie game developer Popcar’s argument that AI-powered decompilation of existing games is now happening at scale, enabling mashups and a boom in preservation projects that get old games running on new hardware, often in browsers via WebAssembly. Popcar frames the upside as better software ownership after servers shut down, but warns the same capability applies to any software: paid games or apps can be decompiled, reused, stolen from, or open-sourced against the owner’s will.

Why: If you ship a paid downloadable app or game, this is a concrete risk to plan around: the text claims AI decompilation makes reverse engineering cheaper and harder to stop, so closed binaries are less of a moat and an unwanted fork is more plausible. For preservation-minded builders, the same trend is an opening to port and keep old software alive via WebAssembly; for founders, decide how much defensibility should come from licensing, services, or community rather than distributed code alone.

Claire Vo reviewed Meta's newly launched Muse Wonder and Muse Wonderful frontier models on the How I AI bench, running them headlessly through Claude Code across nearly 20 tasks covering prototyping, agentic workflows, coding, writing, and creative work. She reports Meta's 'personal and professional agent' framing mostly holds up, with Muse ranking first in three agentic tasks, but design output is not good and it is slow on agent tasks despite being fast on one-shot work. The piece also compares Muse to Sonnet on voice assistant tasks and to Opus 5.5, Astra, and Fable on SVG and short-form video.

Why: If you are choosing models for agent workflows, this suggests Muse Wonder may be worth testing for three agentic task categories, but not as a default for design, SVG, or short-form video work; it also warns that agent tasks may be slower even when one-shot prompts feel fast. Treat it as a first-look signal to run your own eval rather than a settled benchmark, since the excerpt gives no exact scores.

Not-Ship's Amanda Shendruk argues that Amazon's stacked seller fees — not just the 15% referral commission that matches eBay and Walmart, but fulfillment, storage, returns and ads — now take close to half of many sellers' sale proceeds, up from about 19% in 2014. Only the referral fee is formally mandatory, but the piece says Prime eligibility and ad visibility make Fulfillment by Amazon and sponsored placements effectively required, and the FTC's upcoming antitrust case calls Prime eligibility 'a virtual necessity for doing business on Amazon.' Because the referral fee is a percentage of price, higher prices raise Amazon's cut too; the article's next section, on sellers being unable to charge less elsewhere, is cut off in the excerpt. The Hacker News thread drew 701 points and 590 comments.

Why: If you sell or plan to sell on Amazon — including cross-border from Malaysia — model the take rate at roughly 50% of sale price, not the headline 15%, and treat FBA and ad spend as fixed costs rather than options, since the piece's claim is that opting out means not selling. For founders building or benchmarking marketplace software, the useful comparison is take-rate composition: a 15% commission plus fulfillment, storage, returns and ad fees is a different business than a flat 15%, and that gap is what the FTC case will argue about. The excerpt cuts off before the 'can't charge less elsewhere' argument, so don't repeat that claim as fact without reading the full post.

ASML reported third-quarter profit that beat expectations and said some customers are expressing interest in placing orders for its lithography tools for 2029 and even 2030. CEO Christophe Fouquet said short- and medium-term equipment demand forecasts are rising, driven mainly by AI-related investments, and that planned supply for 2027 and 2028 will not fully meet customer demand; ASML is investigating ways to raise 2028 capacity beyond a previously forecast 30% increase. During the quarter, Samsung and TSMC committed to using ASML's High NA EUV lithography machines.

Why: If you build or buy AI infrastructure, this signals that leading-edge chip capacity—and therefore accelerator and cloud GPU supply—may stay tight into the late 2020s. Plan AI compute budgets and model-size roadmaps with longer lead times instead of assuming near-term price relief, and treat Samsung and TSMC's High NA EUV commitments as a concrete indicator that advanced-node AI hardware demand is still expanding. For builders in Malaysia or Southeast Asia tied to semiconductor, electronics, or data-center supply chains, this is a demand signal worth watching for local capacity and services opportunities.

Anthropic announced public beta OpenTelemetry tracing for Claude Managed Agents, the Messages API, and the Claude SDKs, so every agent turn, model call, and tool/MCP server call exports as a standard OTel trace with latency, tokens, and IDs. Traces follow the OpenTelemetry GenAI semantic conventions and can go to any OTLP endpoint, with named integrations including Grafana Cloud, Honeycomb, Sentry, Splunk, Arize AI, Braintrust, Datadog, Langfuse, LangSmith, New Relic, PostHog, and Databricks Unity Gateway. Managed Agents and the Messages API require a destination configured in the Claude Console; the Claude SDKs export traces automatically.

Why: If you already ship Claude-based agents, this is a config change, not a migration: add a destination in the Claude Console (or rely on auto-export from the SDKs) and traces land in whatever OTel tool you already run, sitting next to your logs and metrics so a slow or failed agent step maps back to a service. The concrete decision is whether to turn this on before you widen an agent's tool and MCP access — the post frames tracing as the prerequisite for granting more access, and the Ramp quote claims trace-scoring loops cut token usage 10–20% and made sessions 30% faster on some agents. It is beta, so treat it as instrumentation you can adopt early rather than something to bet a compliance story on.

The Hacker News reports that Docker has fixed a critical flaw in Docker Hub that allowed one user to push images as another user, and that the bug had been present for roughly two years before the fix. The supplied article text contains only the headline and site navigation boilerplate — no CVE identifier, affected versions, patch version, disclosure timeline, or statement from Docker is included.

Why: If you pull public images from Docker Hub, this is the case for pinning by digest (image@sha256:...) rather than by tag or by trusting a namespace, since the reported bug let an account push under another identity. Because the text gives no CVE, no affected version, and no fix version, you cannot act on it beyond that general practice — treat it as a prompt to check your own image provenance controls, not as a patch instruction.

Pluto Security emerged from stealth with $42 million after starting sales in April, crossing $2 million in ARR with 30+ enterprise customers including Oscar Health, Workato, Sixt, and AlphaSense; its $8.5M seed was led by TLV Partners with Armis' founders investing. CEO and co-founder Shahar Bahat says a newly released AI tool can appear inside a company's environment within two hours of launch, and that roughly 60% of the risks Pluto detects across its customer base sit in tools vibe-coded by non-developers. Finance departments are the top source of those risks, followed by marketing and legal, with developers fourth; Pluto cites a Fortune 500 finance employee's vibe-coded customer/churn app left exposed to the internet.

Why: The number to act on is the 60%: if most risky AI-built software in enterprises originates from finance, marketing, and legal rather than engineering, then code review and security gates aimed only at the dev team miss the majority of exposure. The Fortune 500 example is a customer-data app left internet-accessible, so the concrete check is whether internal tools built by non-engineers in your org are publicly reachable and who owns reviewing them. Bahat's two-hour adoption claim also means tool allowlists and blocklists go stale fast, so detection needs to look at what AI connectors and apps are actually running, not just what was approved.

LiquidAI released LFM2.5-24B-A2B, a 24B-parameter MoE reasoning model with only 2B active parameters per token, positioned for running agents on your own hardware. It reports 184 tok/s on an Apple M5 Max and up to 21K output tok/s on a single NVIDIA B300 GPU, and claims 48.09 on LiveClaw versus 38.19 for Qwen3.6-35B-A3B, a model with roughly 50% more total and active parameters. Versus its LFM2-24B-A2B predecessor it adds an explicit chain of thought, grows context from 32K to 131K tokens, doubles tokenizer vocabulary from 65K to 128K for non-Latin text, and scales pre-training from ~17T to ~33T tokens.

Why: If you pay per-token for cloud agent loops, a 2B-active reasoning model means continuous multi-step tool-calling can run on a laptop or a single GPU instead of an API bill — the 131K context and 128K vocabulary are the parts that matter for long tool-use sessions and non-Latin prompts. But the LiveClaw 48.09-vs-38.19 comparison is LiquidAI's own benchmark, so run your own harness eval before replacing a cloud model in anything you ship.

Cloudflare published a deep dive into how the D1 control plane (its serverless SQLite service) was re-architected to support platforms with millions of created databases, hundreds of thousands active at once. The post reports 15x growth in live non-deleted databases over the past year and 65x over two years, plus 18x growth in daily active databases, as of October 8, 2026. It argues that AI-assisted and vibe coding platforms are pushing applications toward one small database per user, app, preview environment, or tenant, and cites ChatGPT Sites at OpenAI using D1 for leaderboards, comments, and custom metrics.

Why: If you're building a multi-tenant SaaS, a preview-environment-per-branch workflow, or a vibe coding platform, this is a datapoint that the database-per-tenant model is being run at real scale on managed SQLite — the growth figures (18x daily active databases in a year) come from a vendor's own numbers, not independent measurement. Note the excerpt cuts off before the actual architectural details, so you can't yet copy the control-plane patterns it promises; treat this as a signal to evaluate per-tenant D1 versus a shared-schema Postgres before your next tenancy decision, not as an implementation guide.

Character.ai, described in its own filings as the startup behind one of the largest AI companion apps, filed for Chapter 11 bankruptcy in the Delaware District bankruptcy court, saying the platform had become "costly to develop and grow" over its 4-year lifespan. The filing cites high costs "coupled with potential litigation related liabilities and regulatory engagement" leaving it unable to generate sufficient cash flows, and notes Kentucky sued the company in January over chatbot safety. Disney has agreed to act as a stalking horse bidder in the sale process, licensing some of Character.ai's IP for $20 million.

Why: The headline number to internalise is $20 million — that is the stalking-horse price Disney put on licensing part of Character.ai's IP, against a filing that blames development cost, litigation liabilities, and regulatory engagement for the cash shortfall. If you are building or planning a companion/roleplay product, treat per-conversation inference cost plus safety/legal exposure as a first-class line item, not an afterthought: a large engaged user base did not save this company. For teams weighing a character or roleplay app as a first AI project, this is a concrete data point that distribution and engagement alone do not cover the cost of running the models.

Cloudflare's Data team describes how its pipeline — Logfwdr into centralized receivers, then the Buftee buffer, then consumers writing to ClickHouse — was handling roughly 2 billion events per second as of May this year, while hardware growth stayed linear. Across five years they made five changes (breaking up the monolithic message schema in 2022, adjusting compression in 2024, rearchitecting pipeline replication in spring 2026, decentralizing sampling in summer 2026, and adding aggregations to the pipeline in fall 2026) that together deferred about two and a half years of hardware expansion. The excerpt only details the 2022 schema split, where a single stream carrying every product's fields was replaced with product-specific schemas that duplicate some fields but stop internal consumers from ingesting everything and filtering it out.

Why: If you run high-volume ingestion into ClickHouse or a similar store, the concrete lever here is schema splitting: stop shipping one firehose message with every product's fields and make consumers demux it, and accept field duplication in exchange for not paying to move and store fields nobody reads. Cloudflare attributes roughly 2.5 years of deferred hardware spend to five such changes at 2B events/sec — but the post only explains the 2022 change in the text available, so treat the 2026 replication, sampling and aggregation items as named but unsubstantiated until the full post is read.

Character.AI, the chatbot platform founded in late 2021 by Noam Shazeer and Daniel De Freitas, filed for Chapter 11 bankruptcy on October 13, 2026, listing assets and liabilities between $50 million and $100 million after what CEO Colin Adams described as legal, regulatory, and operational headwinds. The company has agreed to sell some assets plus a non-exclusive perpetual technology license to Disney for $20 million, with former CEO Karandeep Singh Anand and other former Character.AI staff having joined Disney the previous month. No bidder for the remaining assets has emerged yet; the sale is expected to close in late November or early December, and Adams said the platform will be wound down if no actionable bid appears.

Why: If you or your product depends on Character.AI's platform, characters, or any hosted API surface, treat a late-November/early-December wind-down as a real possibility and plan a data export and migration path now. The $20 million Disney price for assets plus a perpetual license, against $50-100 million in stated assets and liabilities, is a concrete data point on how little consumer chatbot scale converted into exit value five years in — useful when you model your own consumer AI app's terminal value or negotiate with acquirers.

In an Oct 12, 2026 LinkedIn post, Rich Hickey argues that 'tool' should keep a distinguishing meaning — something you wield, held in hand, with power, control, continuous direction and sensory feedback — and that on those criteria LLMs qualify as automation rather than tools, because they are not predictable, reliable, deterministic, or of bounded function. He compares an LLM to a drive-thru window, which is 'automation for avoiding cooking,' and says LLMs are 'automation with which you avoi[d]' writing code (the excerpt cuts off mid-sentence). The piece drew 288 points and 153 comments on Hacker News.

Why: If you ship or buy AI coding assistance, Hickey's list is a checklist of what you are not getting: predictable, reliable, deterministic, bounded-function behaviour with continuous feedback. The practical implication is that verification stays with the human, because 'expertise' in his framing is built by wielding, not by delegating — so review and test gates around generated code aren't optional overhead. For founders pitching an LLM product, calling it a 'tool' invites exactly this objection; the text gives no pricing, version, or benchmark data to settle it either way.

In a Lenny's Newsletter How I AI episode published Oct 12, 2026, Meta design director Alex Cornell describes how Meta designed the Muse AI agent around human conversation rather than software interactions. Key design patterns include progressive disclosure with three transparency levels (emoji, task list, raw logs), a fixed generative-UI structure with dynamic content, and hiding chain-of-thought because Muse is fast enough to operate quietly. Cornell also describes using Muse to turn years of Instagram saves into a shopping catalog, analyze golf and chess games, and transform a 20-year-old blog into a daily podcast.

Why: If you are building an AI agent, this gives concrete UX choices to copy or reject: use tiered transparency so casual users see an emoji while engineers can inspect raw logs, and do not display reasoning just to mask latency if your agent can respond quickly. It also points to Instagram saves as an overlooked personal-data source for consumer agent use cases, which founders can test without building new data capture. No Malaysia or Southeast Asia-specific funding, policy, cloud, telco, payments, or government angle is present in the text.