Zhipu shipped GLM-5.3 via API on August 14 and sat on the weights for two weeks of security hardening, publishing 84.5% on CyberGym to justify the delay. Nvidia is optimizing silicon for Chinese open models regardless.
GLM-5.3 went live by API on August 14, but the weights waited two weeks for cybersecurity review. Zhipu published 84.5% on CyberGym at launch, slightly ahead of Claude Mythos 5 and GPT-5.6 Sol, to argue the delay bought something real.
The stealth OpenRouter model builders have been probing for days is a preview of GLM-5.3 Flash. Artificial Analysis scored it 57 on overall intelligence, level with Claude Opus 4.8, with Z.ai shares up about 10% on the launch.
Customer data was accessed at Manchester, Stansted and East Midlands airports this week. Operators said passenger safety and flight operations were not affected, but the breach lands the same week security gating dominates the model conversation.
Koda publishes a fact-checked AI intelligence briefing every morning: site, email, podcast, and video. Founding sponsors lock today's rates for 6 months as the audience compounds.
View the media kit →Zhipu AI had set August 28 for the GLM-5.3 open weights, two weeks after the model went live via API on August 14, but the zai-org repository was still a placeholder at publication. The weights stay held back for cybersecurity review, and at the August 14 launch Zhipu published 84.5% on CyberGym, slightly ahead of Claude Mythos 5 and GPT-5.6 Sol, plus 54.4% on ExploitBench...
Continue reading arrow_outwardZhipu had still not published the GLM-5.3 open weights two weeks after the API launch on August 14, holding them for a cybersecurity review; at launch the model scored 84.5% on CyberGym and 54.4% on ExploitBench, up from 24.4% for GLM-5.2.
Nepal's Himalayan flash flood death toll reached 362 with nearly 1,400 people still missing, and both China and Nepal warned that two glacial lakes near their shared border could burst.
Mood is Greed: Nvidia's stronger-than-expected revenue forecast pushed the Nasdaq ahead of other major indices, reversing last week's rate-and-Iran-driven tech selloff.
Nvidia is simultaneously optimizing its hardware and memory stack for DeepSeek and Qwen, warning the SEC about possible Trump administration restrictions on China-developed models, and committing with AWS to as many as 2 million GPUs across 2027 and 2028: three bets that do not obviously point the same way.
Zhipu shipped GLM-5.3 by API on August 14 but withheld the weights for two weeks of cybersecurity hardening, and they had still not appeared at publication. It published scores at launch to justify the delay: 84.5% on CyberGym, slightly ahead of Claude Mythos 5 and GPT-5.6 Sol. Release timing is becoming a safety lever rather than a marketing one, and rivals now have a template for staggered launches.
The stealth OpenRouter model tracked for days is a preview of GLM-5.3 Flash, scored 57 on Artificial Analysis overall intelligence, level with Claude Opus 4.8. Z.ai says it led OpenRouter usage while running on 100,000 Chinese-made AI chips, and its shares rose about 10%, tying domestic silicon directly to a market-visible result.
AWS and Nvidia outlined an expansion of up to 2 million GPUs across 2027 and 2028 with no capital figure disclosed, extending the buildout two years past capacity already under construction. Nvidia's revenue forecast lifted the Nasdaq, but the same rate pressure that drove last week's selloff still sits underneath these unpriced commitments.
17,600 Attacks From Its Own Agents
Read full analysis arrow_outward
Deep Dive
6.6/ 10
It pitches itself as the agent that saves agencies $60K a year, but the loudest verified signal is 11 Product Hunt reviews and unpublished paid pricing.
KIVA is for small agencies and solo marketers who already have Google Search Console data and want an agent to convert it into keyword clusters, briefs, and drafts aimed at both classic search and AI answer engines. The free-forever tier makes it cheap to validate on one site, and the 2.0 shipping cadence suggests real momentum. But public evidence is thin, third-party ratings swing wildly from 4.6 down to 2.8, and paid pricing is unpublished, so treat it as a trial rather than a stack replacement.
“A reviewer liked being able to pick the LSIs and People Also Ask questions that fit their needs, then drop the resulting content straight into their blog.”Product Hunt review
AWS expanded Quick's Microsoft 365 integration so the agent operates in the files you already have open: spreadsheet analysis, slide drafting, tracked edits in Word, plus inbox triage and meeting scheduling in Outlook. Pilot it on one repetitive artifact, like a weekly deck or a recurring variance analysis, and check whether tracked changes are reviewable before you let it touch shared documents.
Try it arrow_outwardZoom added agentic automation for voicemail and post-call processing: follow-up checks, forwarding, Team Chat and SMS notifications, email, and ticket creation in external systems. The useful pattern here is not the transcription, it is the handoff, so wire it to one queue where dropped callbacks actually cost you money and measure the change in response time.
Try it arrow_outwardGoogle's voice assistant now runs agentic Spark tasks, reads a spoken Daily Brief, controls the Gmail inbox, and pulls Personal Intelligence across connected apps, meaning a spoken request can trigger edits in Docs, Sheets, and Drive. Before enabling it broadly, decide which apps it may read, since the value and the exposure both come from the same connected-account access.
Try it arrow_outwardOpenAI's office-focused agent platform is pitched at complex workplace tasks with a simpler interface, and the reported early wins are unglamorous: calendar migration, auto-updating dashboards, recurring reports. Those are the right test cases because they have a verifiable output, so pick one and compare the agent's result against the version a person already produces.
Try it arrow_outwardAudit your Copilot model policy today. GitHub made its global model policy public and open-weight models including DeepSeek and Kimi K2 are default-disabled in some enterprise configurations. Check whether your org is silently blocking the models your benchmarks assume, then decide deliberately rather than by default.
Pilot Amazon Quick inside Office. AWS expanded Quick's Microsoft 365 integration so the agent operates inside Excel, Word, PowerPoint and Outlook. Pick one recurring report and let the agent draft it in place instead of exporting data to a separate chat window.
Benchmark GLM-5.3 Flash before you commit. The model formerly known as Ox Alpha scored 57 on Artificial Analysis overall intelligence, level with Claude Opus 4.8. Run it against your own eval set on OpenRouter this week and compare cost per solved task, not leaderboard rank.
Zhipu's two-week weight embargo is the first real test of whether a lab can turn a delay into a credibility asset, and the 84.5% CyberGym number is the receipt it chose to publish. Meanwhile Nvidia is tuning memory stacks for DeepSeek and Qwen, and GitHub is default-disabling those same models in enterprise Copilot. The infrastructure layer wants Chinese open weights and the distribution layer does not, which means your model access is now a procurement decision as much as a technical one. Check what your tooling actually permits before you architect around a model you cannot ship.
AI analysis, world news, markets, and tools. One briefing, delivered free.