AI news5
Microsoft patches "CoSnitch," a critical Copilot flaw that let one click silently exfiltrate a victim's Gmail, Drive and OneDrive data
Microsoft shipped a complete fix on 18 August for CoSnitch (CVE-2026-24301), a critical Copilot Personal flaw Varonis Threat Labs reported in December 2025. Three chained weaknesses let a single clicked link auto-execute an attacker prompt, pull data from a victim's connected Gmail, Drive or OneDrive accounts, exfiltrate it via Copilot's own URL-fetch capability, and poison the assistant's permanent memory store in a way that survived password changes and device re-enrollment. Microsoft found no evidence of active exploitation before the patch, roughly 7.5 months after the initial report.
csoonline.com ↗Why it matters
a single-click attack chain against a mainstream AI assistant's connected accounts, unpatched for seven months, previews the attack surface every AI copilot with account access now carries.
China's LandSpace lands an orbital-class rocket booster on legs for the first time
LandSpace recovered the first stage of its Zhuque-3 rocket on 18-19 August in Gansu province, 390km from the launch pad, becoming the first Chinese private company (and only the third globally after SpaceX and Blue Origin) to land an orbital-class booster on legs. LandSpace plans to reuse the booster within six months.
techstartups.com ↗Why it matters
reusable boosters are the biggest lever on launch cost, and China closing this gap with SpaceX changes the competitive picture for commercial space.
Cerebras unveils CS-4, a 750-petaflop rack-scale AI accelerator built from three new wafer-scale chips
Cerebras announced the CS-4 on 18 August: a rack-scale system built from three new WSE-3 Turbo wafers (4 trillion transistors, 900,000 AI cores each) delivering 750 petaflops and 7.2 terabits/second of I/O. On GPT-OSS-120B the company claims up to 30x more tokens-per-second-per-user than leading GPU-based solutions, with first shipments this quarter.
hpcwire.com ↗Why it matters
a 30x inference-speed claim against GPU incumbents, if it holds up independently, directly challenges Nvidia's grip on AI inference, not just training.
China begins receiving limited Nvidia H200 shipments as Beijing eases restrictions
ByteDance and Tencent have each received roughly 10,000 Nvidia H200 processors after Beijing eased its own restrictions on the chips, per Financial Times reporting from 18-19 August corroborated by multiple outlets. Regulators are still steering much of that compute to stay outside the mainland, with some deployment routed through Hong Kong.
tomshardware.com ↗Why it matters
compute has been the binding constraint on China's AI industry, and this is the clearest sign yet that both US and Chinese policy are loosening around the chip that matters most.
Z.ai delays public release of GLM-5.3 weights over hacking-capability concerns
Z.ai shipped GLM-5.3, a 743-billion-parameter open-weight coding model, on 14 August, claiming it beats comparably-sized open models on code benchmarks. The company is delaying the public release of downloadable weights by two weeks specifically because the model is unusually capable at finding and exploiting security flaws, while keeping the hosted version live now.
axios.com ↗Why it matters
a Chinese lab delaying an open-weight release over its own offensive-security capability inverts the usual "China ships fast, safety comes later" framing.