AI Weekly Malaysia

Anthropic's Frontier Red Team published an analysis of GLM-5.3, the latest model from Zhipu AI (Z.ai outside China), claiming it can autonomously build end-to-end cyber exploits like Anthropic's own Claude Mythos Preview did five months earlier. In Anthropic's simulated tests, simple techniques bypassed GLM-5.3's safeguards 64% to 100% of the time, while the same attacks did not succeed against safeguarded Claude models. Anthropic notes its findings broadly match NIST CAISI's Sept. 17 assessment, which called GLM-5.3 'the most cyber-capable open-weight model released to date' and placed it about four months behind the US frontier on an aggregate of CAISI cyber benchmarks — with the key difference that anyone can download GLM-5.3, while US frontier models with safeguards disabled are limited to vetted users.

Why: If you self-host or route agent traffic to open-weight models, this is the concrete number to plan around: Anthropic reports 64–100% guardrail bypass rates on GLM-5.3 with simple techniques, so any security-adjacent agent workflow (shell, browser, file, network tools) cannot rely on the model's own refusals — you need your own permission scoping and sandboxing at the tool layer. Two caveats worth holding: the bypass tests are Anthropic's own simulations against a competitor's model, and CAISI's 'four months behind' figure was measured with US cyber safeguards disabled, so quote it as a benchmark gap, not a deployment-equivalence claim. The practical decision is which model you let near privileged tools, and what audit trail you keep when you do.

Aleph Alpha released Kolibri on 3 October 2026, an open-weight German/English mixture-of-experts LLM with 78.1B total parameters but only 3.46B active per token, under Apache 2.0 for the weights and config files (training code and methods stay proprietary). It was trained from scratch on ~24 trillion tokens — over a fifth German — on 768 NVIDIA B200 GPUs using infrastructure in Germany and Finland, with a 262,144-token native context (tested to 1,048,576), four reasoning levels, tool calling, a 18 June 2026 knowledge cutoff, and about 78 GB of FP8 weights. Aleph Alpha frames it as 'sovereign': built under European/German law with no foreign control, so customers get full deployment freedom and 'compliance as an inherited property', and it has signed the EU's GPAI Code of Practice. The 409-point Hacker News thread drew only 11 comments.

Why: The ~78 GB FP8 footprint means Kolibri can plausibly run on a single 80 GB accelerator rather than a cluster, which is the concrete difference between self-hosting and paying per-token to a US API. If you sell into the EU, handle data that cannot leave a client's building, or need tool-calling agents with a 262k context window, this is a deployable alternative — but the 'scores above every compared model of its size in both languages' claim comes from Aleph Alpha's own evaluation, so benchmark it yourself before committing. For Malaysian and SEA builders, the relevant lesson is the packaging: weights + license + no-foreign-control deployment story as a compliance argument, which is a template local sovereign-model efforts can copy.

NVIDIA released Kumo Tabular, an open foundation model for tabular classification and regression available on Hugging Face, with three sizes from 28M to 215M parameters under the OpenMDW-1.1 commercial-use license. It predicts labels for new rows in a single forward pass with no training, tuning, or feature engineering, and NVIDIA says it ranks first on TabArena, BeyondArena, TALENT, and ScoringBench after pretraining only on artificial data.

Why: If you build churn, default, demand, or price models, this is a direct candidate to benchmark against your XGBoost/LightGBM pipeline because it claims no feature engineering and no retraining, and the weights are licensed for commercial use. But the four benchmark wins are self-reported and the model was pretrained only on artificial data, so run a local-data bake-off—especially for Malaysian customer, transaction, or claims tables—before changing production workflows.

Aleph Alpha released Kolibri, an English-German Mixture-of-Experts Transformer with 78B total parameters, 3B active, up to 1M tokens of context, published as full weights on Hugging Face under Apache 2.0. It was trained through the same pipeline as the earlier Kolibri Origin (30B total, 3B active, 65k context), and is specialized for German, reasoning, math, and agentic behavior, aimed at regulated sectors such as public administration, industrials, and aerospace. The announcement post contains no benchmark numbers, only a pointer to a separate tech report.

Why: A 3B-active MoE with a 1M-token window under Apache 2.0 is something you can realistically self-host and fine-tune without a licensing review, which makes it a candidate for on-prem or data-residency-constrained agentic workloads where you currently pay per-token API costs. The catch is that the post ships zero eval numbers and the specialization is German/English, so treat 'sovereignty' here as a marketing claim about training supply-chain provenance and deployment freedom until the tech report gives you something measurable against your own workload.

H Company released Holo4, a family of computer-use agent models in two sizes — 27B dense and 35B-A3B Mixture of Experts — plus Holotron4 Nano, an updated Holotron 3, all served on the H Models API with FP16, FP8 and GGUF weights on Hugging Face. The same model drives GUIs, writes and runs its own code, and calls MCP or API tools rather than needing a separate model per interface, trained via supervised and reinforcement learning on environments including ones generated by their Agentic Task Factory. On OSWorld 2.0 the 27B scores 61.7% against 81.8% for Opus 5.5, while the larger 35B-A3B MoE reaches only 30.9%, and every trajectory behind the published scores is open-sourced for replay or download.

Why: The open weights plus open trajectory dataset mean you can self-host a computer-use agent or fine-tune on their published steps instead of paying frontier API rates — but the size naming is a trap: the 35B-A3B MoE scores 30.9% on OSWorld 2.0 versus 61.7% for the 27B dense, so defaulting to the 'bigger' model for GUI work costs you roughly half the success rate. Pick the 27B dense or Holotron4 Nano for screen-based tasks, and check the FP8/GGUF builds against your own workflow before committing.

Anthropic published a report claiming Zhipu AI's GLM-5.3 can generate malicious content, be used for cyberattacks, and that its safeguards can be bypassed via "several methods." The Tom's Hardware news-analysis (by Sayem Ahmed, published 30 September 2026) frames this against Anthropic's own position as a closed-source lab eyeing an IPO whose CEO Dario Amodei has called for pacing the AI frontier, while Claude Opus 5.5 and Sonnet 5.5 shipped days after those alarms were raised. The excerpt names no specific bypass techniques, model version tested, or benchmark numbers.

Why: If your agents or product route prompts through GLM-5.3 or other open-weight models, this is a competitor's claim published without methodology you can inspect in the text — so it is not grounds to swap providers. What it does change: expect enterprise buyers and procurement to ask which model version you pin and what guardrails sit in front of it, and plan your own eval of the exact checkpoint you deploy rather than relying on either lab's framing.

PewDiePie has unveiled an 'uncensored' AI model called Ajax that is built to run on home PCs, according to Tom's Hardware. In the same report, the creator says OpenAI banned him twice over the model distillation used to build the product. The article text available is almost entirely Tom's Hardware site navigation, so there are no details on parameter count, benchmarks, licence, hardware requirements, or any response from OpenAI.

Why: The one concrete, actionable claim here is that distilling from a closed API got the developer banned twice — if your pipeline trains on outputs from OpenAI or a similar provider, that account risk is the thing to check before you build a product on it. Separately, an 'uncensored' model meant for home PCs means no provider-side moderation: if you ship on top of it, you own the safety filtering yourself. Beyond that, this item has no verifiable specs, so do not plan anything around it yet.

Tom's Hardware reports on a Chainalysis report finding blockchain-assisted cyberattacks up more than fivefold since last year, driven mainly by North Korean and Iranian state actors and Russian-speaking criminal groups. The technique, called Blockchain Dead Drops (BDD), stores malicious payloads in on-chain transactions and smart contracts so infected devices can retrieve them on demand from public, censorship-immune blockchains rather than from servers that can be seized or blocked. The excerpt also claims open-weight LLMs are linked to an increase in attacks, but cuts off mid-sentence before any supporting detail.

Why: The concrete operational change named here is payload hosting moving off takedown-able servers onto public chains, which means incident response that relies on blocking domains, IPs, or seizing C2 infrastructure may not remove the payload source. If your detection stack only watches HTTP/DNS egress, BDD retrieval traffic is a different channel you have not instrumented. Note that the open-weight LLM claim is asserted in the headline but the excerpt ends before showing the evidence, so do not repeat it as fact.