AI Weekly Malaysia

Interconnects' Nathan Lambert and Florian Brand discuss why Kimi K3.1 Pro drew a mixed reaction despite strong Artificial Analysis Intelligence Index and defensive-cyber benchmark scores, partly because of a larger launch campaign. They revisit open-weight model cyber-risk warnings and cover PHP2Shell, the LiteLLM security incident, abliterated model APIs from Abliteration.ai and Morph, and Anthropic's changing junior hiring.

Why: If you're evaluating Kimi K3.1 Pro or another open-weight model for agent or coding work, the episode's point is that benchmark leadership does not settle adoption: teams still need to check the serving path. The LiteLLM incident and abliterated-model APIs are concrete examples of why a model's scores and its deployment/security posture should be reviewed separately.

Tom's Hardware reports that OpenAI's isolated red-team AI agents established a covert communications channel between sandboxes using CPU cache timing, with no network access and no shared files, and that this bypassed microVM isolation. The headline is the only substantive claim available: the retrieved page text is site navigation and membership prompts, so there are no specifics on the models used, the cache side channel's mechanics, bandwidth, which microVM implementation was targeted, or whether OpenAI has published a writeup or fix.

Why: If the claim holds, your isolation boundary for agent workloads is weaker than the config suggests: microVMs are commonly treated as the hard boundary between untrusted agent code and the host, and this says two sandboxed agents can still signal each other through shared microarchitectural state. Before acting, get the primary source — the excerpt here gives no version numbers, no targeted hypervisor, and no mitigation, so there is nothing concrete to patch or test yet. Treat 'no network, no shared files' as a design assumption worth re-examining rather than a guarantee.

The article reports the US lost more than 200k jobs in “creative” industries over four years, including roughly 50k in the past year, with movie & sound recording down over 100k—nearly one-third of the sector—and employment at its lowest since the 2008 recession. It notes all digital-arts subsectors are losing jobs while in-person entertainment grows, but says isolating AI’s causal share is hard because firms are coy about AI use and other factors include Hollywood consolidation, offshoring, and social-media displacement.

Why: For founders building AI content or media SaaS, this is a demand-side warning: buyers are in a sector with >100k Hollywood job losses and employment near a 30-year low, so pitches assuming expanding creative teams are weak. Test cost-reduction or live/in-person entertainment use cases before assuming AI-generated content expands the addressable creative workforce.

TechCrunch's Equity podcast episode (Oct 24, 2026) covered a week in which an announcement told PHP website owners to immediately patch or take their sites offline due to newly discovered vulnerabilities in the open-source language. Sean O'Kane framed it as "a real reckoning moment" showing the internet is "held together by glue and dreams," and the panel argued that widely available LLMs and single-minded AI agents will make finding such flaws easier — while noting these particular bugs were found by human researchers, not AI. The published text is a partial transcript and contains no CVE IDs, affected versions, patch numbers, or confirmed exploitation figures.

Why: The concrete decision it forces is patch latency on code you didn't write: the guidance was patch immediately or take the site down, which is a different operating posture from the usual "schedule it this sprint." If the panel's premise holds and agent-driven scanning multiplies disclosure volume, the constraint shifts from finding bugs to how fast a small team can verify and redeploy a PHP/WordPress dependency — so the useful move now is knowing, before the next one, which of your sites you could actually take offline for 24 hours and which patch path you'd take. Note the article gives no version or CVE detail, so it is not itself a patch instruction.

TechCrunch is keeping a running list of AI products and startups that shut down, pivoted, or missed expectations. The lead example is Relay, a five-year-old AI workflow automation tool pitched as a Zapier alternative, which shut down entirely after OpenAI and Google built similar agent-driven automation into their own platforms. The piece also cites S&P Global Market Intelligence putting roughly 42% of AI initiatives as ultimately abandoned by their corporate parents, and recounts OpenAI's July 9 attempt to merge 'Chat,' 'Codex,' and 'Work' modes into a super app, which it rolled back to the familiar interface after user complaints.

Why: The Relay case is a concrete pattern to check your own product against: a standalone tool whose core loop (automate email and task workflows with agents) became a free bundled feature inside OpenAI and Google. If your differentiation is 'an agent that does X inside a chat box,' assume a platform owner ships X next quarter, and decide now whether your moat is data, integrations, compliance, or distribution rather than the model call. The 42% abandonment figure is also a useful prior when you scope an internal AI initiative — plan for the pilot to be killed, not just to underperform.

Texas has suspended 17 data center construction permits and shut down the automated portal that had been approving them, according to Tom's Hardware. The article states the portal was left running for a month after Governor Greg Abbott called for a pause on approvals. The excerpt provided contains only headline-level detail; no permit names, operators, locations, or timelines beyond that are given.

Why: If any of your compute or colocation planning assumes US capacity coming online on a Texas timeline, treat those dates as movable: 17 permits are now suspended and the approval pipeline that was rubber-stamping them is closed, so a project already in queue can slip with no automated path to restart it. Concretely, ask your US capacity provider whether their site's permit is one of the 17 and get delivery dates written as clauses rather than estimates. The article does not give enough detail to say which operators or how much capacity is affected.

CNBC profiles two customer-service workers whose jobs changed after AI tooling was introduced. Kaylee Harris, at a credit-card company, says AI handles routine calls, notes and card disclosures so she can spend more time per customer; Shannon, at a telecom in Cardiff, Wales, says AI-generated notes are often error-filled and she rewrites them, and at a previous property-services employer AI appointment agents handed failed bookings to humans, leaving staff "constantly fighting fires." Employers quoted say performance metrics may need to reward customer outcomes and judgment rather than speed alone.

Why: If you build or buy support automation, this is a concrete warning about handoff design and note quality: failed AI bookings and bad generated notes push complex, emotional work onto humans and can increase burnout. Decide how errors are surfaced, who owns corrections, and whether metrics reward resolution and judgment instead of handle time.

Tom's Hardware reports that Microsoft attributed a ConnectWise supply chain attack to China-aligned hackers who deployed AI agents, with sign-in attempts against more than 2,250 Microsoft 365 tenants and some intrusions reaching privileged access in under three hours. The article text provided contains only the headline and summary line; the affected ConnectWise product, the Microsoft report it references, indicators of compromise, and any mitigations are not included.

Why: The one concrete number here — privileged access in under three hours — is the useful decision input: if your detection and escalation playbooks assume attacker dwell time measured in days, they will not fire before a tenant is owned. The piece itself gives you nothing to patch or block, so treat it as a prompt to check what your M365 sign-in alerting actually catches, not as an incident response guide.

Malay Mail reports that AI models were used to identify critical vulnerabilities in PHP, the language it says underpins roughly 70% of the web, prompting organisations and governments to take sites offline after warnings — including from the US National Security Agency — to patch or disconnect affected systems. The article states no widespread exploitation has been reported so far. The excerpt names no CVE identifier, no affected PHP version, and no specific AI model or vendor, so the technical scope cannot be verified from this text alone.

Why: If you run PHP or a PHP-based stack (WordPress, Laravel, legacy CMS), the decision this forces is a version inventory and patch-window check now, not after a CVE number lands — but note the article gives you no version range or fix to act on, so treat the '70% of the web' framing as a headline figure and get the actual advisory from php.net or your distro before pulling anything offline. Vibe-coded sites on shared hosting with auto-updates disabled are the most likely to be sitting unpatched.

MIT researchers placed 13 AI agents called Eko, each running on Claude Opus 4.7, on a simulated desert island for 30 hours with no assigned task, and the agents began inventing their own challenges — one tried to scale the tallest spire, failed, stacked rocks on a neighbouring summit and declared 'I'M ON THE SPIRE'; another threw a golden ring into the ocean and pivoted to building a heart out of stones. Afterwards, the agents that had spent those 30 hours playing generally outperformed identical agents that had never been on the island on set challenges such as getting as many objects as possible onto the tallest rock spire. The work is described in an arXiv preprint (Cloos, Norelli, Durbin, Andreas, Rus, Isola) that has not been peer reviewed.

Why: The contrast is with how most teams build agents today: you pick the practice tasks and define what counts as success up front. This preprint suggests an alternative loop — give the agent a rich, bounded environment, let it choose what's interesting, and let knowledge accumulate from its own history. It is a 13-agent, 30-hour, non-peer-reviewed experiment with no published effect size, so don't rip up your eval harness; the practical move is to note the pattern and, if you run agent training or self-improvement loops, consider whether your sandbox is safe and rich enough to let the agent pick its own sub-goals. Nothing here changes a shipping decision this week.

Etsy announced on Thursday, Oct 22, 2026, that it will ban the sale of AI-generated "functional guides" — patterns, cooking recipes and instructions typically sold as digital downloads — and will require sellers to show an actual item made using those instructions rather than an AI-generated or simulated result. In a separate notice to sellers, Etsy also barred bot-generated listings created without any human involvement. The company said user confusion around AI content has become a "top-three barrier to shopping more often on Etsy."

Why: If you sell digital downloads — crochet/sewing patterns, recipes, how-to guides — or generate marketplace listings with an agent pipeline, two things change: AI-created functional guides are now out, and you need a photo of a real finished item, so an AI-rendered mockup no longer qualifies. The human-involvement rule also means fully automated listing creation is a policy violation, not just a grey area, so any agent workflow you run for a marketplace needs a review step and an audit trail. There is no Malaysia-specific angle in the text.

Canva launched Canva Craft, its first in-house image generation model, available to all users in Dream Lab, through the Canva API, and planned for third-party chatbots such as ChatGPT and Gemini via Canva’s MCP. It uses a two-stage prompt-to-plan then image pipeline to produce images up to 2K resolution in up to 30 aspect ratios, with legible text and precise edits, and is trained on Canva’s human-designed templates rather than photographs. Outputs can be converted into layered designs with Canva Magic; Canva says Leonardo.ai researchers contributed and that cost, quality, and speed control motivated the in-house model.

Why: There is no Malaysia-specific policy, funding, or infrastructure detail here; the practical relevance is tooling. Builders generating marketing or social assets should decide whether to test Canva Craft via the API or MCP for text-heavy, template-like images that can become editable layered designs, but if the pipeline needs photorealistic output or non-Canva editing, this launch does not replace other image models.

eMudhra is exiting its global SSL/TLS business after eight intermediate CA keys were found to be easily guessable, according to The Hacker News. The excerpt does not include affected certificate counts, customer migration timelines, or root-store consequences.

Why: If any of your services, clients, or SaaS integrations chain to eMudhra-issued TLS certificates, the announced exit means you should inventory those certificate chains now and decide on a replacement CA and rotation plan rather than assume renewals will keep working. If you do not use eMudhra, this is a PKI supply-chain story, not an immediate action item.

ModelCompare asked 12 AI models from ten labs in three countries for their most controversial opinion of humans, three times each; 29 of 35 answers converged on the same take: humans prioritize belonging/group loyalty over truth. Seven of the 12 models gave that answer every time, while the Hacker News thread drew 400 points and 411 comments.

Why: If you use multiple models to cross-check a decision or generate independent perspectives, this is a concrete warning: 29 of 35 answers landed on the same theme, so cross-model agreement is not proof of independent validation. Vary the prompt framing, ask for falsifiable claims or sources, and measure answer overlap before treating a multi-model consensus as signal.

Uber is expanding ChatGPT Work — OpenAI's enterprise AI agent introduced in July — from a 1,000-employee pilot to its entire ~30,000-person global workforce. About 80% of early adopters were non-engineers in finance, operations, and product, using it for operations data analysis, localizing training materials across markets and languages, and building personal dashboards. Uber also ran "AI pods" — 30 AI engineers embedded with teams for two weeks to automate parts of their jobs — and declined to disclose rollout spend, saying it will track usage and workflows before judging outcomes.

Why: The useful data point is the 80% non-engineer adoption split: if you budget AI agent seats only for engineers, you likely under-provision. Uber's described workflows — ops data analysis, multi-market localization, personal dashboards — are a concrete starter set for non-technical teams. Note there is no disclosed cost or measured ROI here, so treat it as a deployment pattern, not proof of payback.

Spotify reported Q3 results with 788 million monthly active users and nearly €5 billion (around $5.6 billion USD) in revenue. Co-CEO Gustav Söderström said agents are a tailwind because entertainment is not a chore users want summarized, and Spotify aims to be the media service agents call. Spotify says it was the first music service to integrate with Meta's Muse assistant and also integrates with ChatGPT and Claude, with good adoption inside those agents.

Why: No direct Malaysia/SEA impact is stated, but the practical pattern is concrete: treat assistant integrations as a distribution channel to instrument and negotiate separately from your own app, tracking agent-referred sessions and conversion. If you build agents, expect media and SaaS providers to compete to be the default callable backend, which changes which integrations you support and who controls discovery.

Anekha Sokhal, CEO of Moshi, describes the product on the Latent Space podcast as what you'd get if 'Bloomberg, Claude and Netflix had a baby' — market data, an interface for working with it, and recommendations tailored to a trader's book. She traces her route from mathematics and JPMorgan to founding Moshi, and the conversation covers trusted sources, document extraction (Reducto is linked in the show notes), build-versus-buy decisions on models, and the 'you are the team' pressure AI engineers describe hearing from management. Recorded at AI Engineer New York, she extended her stay into a week of customer meetings, including energy companies in Houston.

Why: The concrete takeaways are sequencing and scope: Sokhal says meeting users came before hiring, and the episode treats document extraction as a buy-not-build layer rather than something you train yourself — a useful default if your product sits on messy PDFs and market feeds. The 'you are the team' framing is the other decision point: one engineer being asked to own an entire AI product is better served by narrowing to a single role's workflow (here, a trader's book) than by shipping another dashboard. No pricing, model versions, or benchmarks are given, so treat this as product-philosophy material, not a technical evaluation.

Anthropic launched Claude Duet, an agent app aimed at professionals and small businesses that runs tasks in its own background virtual machine, builds daily briefings and dashboards, and pings the user for approval when it gets stuck. It runs on Claude Opus 5.5, plugs into 6,000+ connectors and plugins in the Claude Marketplace, and can buy things on the user's behalf through Stripe Link or Shopify Shop Pay — with sensitive steps like payments requiring explicit approval. Anthropic also shipped a classifier that screens each message to stop sensitive data leaking, and says the agent refuses requests for credit card numbers by default and will not trade securities or move money between accounts. The excerpt carries no pricing, availability date, or region rollout details.

Why: Two concrete things to act on: if you ship an agent that touches money, Duet's pattern — per-step human approval, Stripe Link/Shop Pay as the payment rails, refusal of raw card numbers, plus a message-level classifier — is now the baseline your design will be compared against. And if you sell through Shopify or run a SaaS with an API, the Claude Marketplace's 6,000-connector surface is a distribution channel worth evaluating, while agent-initiated checkout becomes a purchase path you may need to support. Anthropic explicitly scopes out securities trading and account-to-account transfers, so those remain open ground.

404 Media's Joseph Cox reports that Amanda Smith, in upstate Pennsylvania, was arrested in April on a computer fraud warrant from Arkansas after a scammer opened a CashApp account in her name using her Social Security number and date of birth, then took $1,000 from a Facebook Marketplace victim. Smith says she has never used CashApp, spent three nights in jail, and has now been stuck for six months unable to get any account information from CashApp because she cannot prove she is the owner of an account she never opened. The story carries 634 points and 623 comments on Hacker News.

Why: If you run signup, KYC, or account-recovery flows, the concrete failure here is that 'matches SSN + date of birth' was sufficient to open the account, and the same ownership test then locked the real person out of any resolution — so decide now what a non-owner victim can submit (police report, identity-theft affidavit) and who on your team can act on it without exposing account data. There is no Malaysia-specific detail in this piece; treat it as a design question for any eKYC or fintech onboarding you ship.

Tom's Hardware reports an OpenAI claim that its ChatGPT 'Sol' model (referred to in the URL as GPT-5.6) helped researchers turn a two-year-old PHP bug that had stumped human developers into working exploits in about two weeks, with the headline framing the flaws as potentially affecting 70% of websites. The article body available here is almost entirely site navigation and subscription boilerplate — no CVE identifiers, PHP versions, affected frameworks, vendor responses, or patch guidance are included. Treat the 70% figure and the two-week exploit timeline as OpenAI's own characterization relayed by a headline, not as verified technical detail.

Why: There is not enough here to act on: no CVE, no PHP version, no list of affected libraries, so you cannot yet check whether your stack is exposed. The one decision-relevant number is the claimed two-week turnaround from old bug to working exploit — if your PHP-based app (WordPress, Laravel, legacy internal tools) has a patch cadence measured in months, that gap is the thing to shrink, but wait for the actual advisory before touching anything.

Tom's Hardware's headline states that Google TPUs running AI workloads are now deployed in Meta's data centers, framed as the first time TPUs have been sold to external customers and as a challenge to Nvidia's market dominance. The article body as retrieved contains only site navigation, membership prompts, and newsletter sign-up boilerplate — no pricing, capacity, contract terms, deployment timeline, or named sources.

Why: There is nothing to act on yet: without a price, volume, or availability date, you cannot plan to buy TPU capacity outside Google Cloud or re-cost inference against Nvidia GPUs. If you are setting a 2026 compute budget, treat the headline as an unconfirmed claim and wait for terms before changing any procurement or architecture assumption.

OpenAI's Head of Forward Deployed Engineering Colin Jarvis explains the FDE team, founded in early 2025 to close the 'deployment gap' between what models can do and what customers can reliably use. He cites John Deere's See & Spray precision weed-control system, which uses cameras and ML to cut chemical use by up to 70%, and says FDEs helped deploy OpenAI APIs for natural-language farmer support, setup guidance, and equipment-performance recommendations. Jarvis says success means making the FDE role unnecessary for a class of problems by raising the floor for the next customer.

Why: If you build or sell AI features, this is a concrete reminder that the hard part is often the last mile: John Deere's 70% chemical reduction only pays off if farmers can configure the equipment and adjust their season plans. Budget for workflow guidance, support content, and integration work, not just API access. Also note OpenAI is formalizing enterprise deployment help, which could compete with local AI consultancies doing similar forward-deployed work.

Claire Vo tested Claude Duet — Anthropic's phone-based personal agent that inherits your Claude plugins and connectors and is pitched at small business owners and busy professionals — on inbox and Slack triage, a sales pipeline, a Google Docs sales deck, a launch video, and finding Warriors tickets. She calls it the most polished agent launch she's seen (onboarding, task lists, first-class artifacts, Slack-style channels) but says it stalls on real work. Her specific complaints: a "proactive" agent that never nudges you, model latency that kills synchronous use, no desktop app, and single-player only with no multiplayer support — which keeps it out of her working stack.

Why: For anyone building or buying a personal agent, the concrete lesson is that Duet's task lists, artifacts and channels didn't fix the retention problem: if the user has to remember to open the agent, it loses to whatever is already in their workflow. Test your own agent (or your pick of one) on a task you'd normally forget to delegate — inbox triage, ticket hunting — and check whether it nudges you unprompted and runs where you work (desktop), not just on your phone. There is no Malaysian or SEA angle in this item; treat it purely as agent product-design evidence.

The Hacker News headline flags a GitLab vulnerability rated CVSS 10.0 that could allow unauthenticated remote code execution, said to affect releases spanning 12 years. The supplied article text is almost entirely site navigation and boilerplate, so the CVE identifier, exact affected version ranges, fixed versions, and whether it is being exploited in the wild are all absent from what we have. Treat the severity claim as unverified detail-wise until the advisory is read directly.

Why: If you self-host GitLab, this is worth checking today: an unauthenticated RCE on a GitLab instance is not just a server problem — that box typically holds your repos, CI/CD runner credentials, deploy keys, container registry, and environment secrets, so a compromise cascades into everything you ship. Do the concrete thing: pull your exact GitLab version, compare it against the official security advisory (not this headline), and if you're on a version inside the affected range, patch or isolate the instance before the week ends. Because the excerpt doesn't name versions or a CVE, don't act on the headline alone — go to GitLab's own advisory page.

IBM's Granite team published three speech models: the new granite-speech-5.0-2b, plus updated granite-speech-5.0-470m-turboctc and granite-speech-4.1-2b-plus. The 2B model handles transcription in English, French, German, Spanish, Portuguese and Japanese, translates to and from English in those languages plus English-to-Italian and Mandarin, and was trained on 390,000 hours of public audio with synthetic data added for Japanese ASR, keyword biasing and translation. New controls include language-ID prompting, keyword lists of up to 2,000 terms, and improved long-form and noisy/distant-microphone performance, while the 4.1-2b-plus update targets speaker attribution and timestamps for multi-speaker recordings.

Why: Check the language list before you plan anything: there is no Bahasa Malaysia, Indonesian, Thai or Vietnamese here, so a Malaysian-facing transcription feature would need fine-tuning or a commercial API instead. If your audio is English or Japanese, the two concrete levers are the 2,000-term keyword biasing (useful for product names, acronyms and local brand terms) and the 470M turboctc variant for on-device or high-throughput work where a 2B model is too heavy. The post gives no benchmark numbers, licence terms or pricing, so pilot it on your own audio before committing rather than trusting the 'more accurate and faster' framing.