Week so far: inference isolation, inherited trust, and staffing decisions

Where LLM serving breaks isolation and accuracy

Cloudflare published a detailed account of a cross-request KV cache exposure in Workers AI and its remediation. KV cache reuse is a standard way to raise LLM serving throughput, so the case shows that tenant isolation in shared inference can fail inside performance code. Cache keying and scoping deserve the review normally reserved for security code. Ubicloud's account of diagnosing tool-call accuracy problems in vLLM points to the same layer for reliability. Parsing and serving details affect whether tool calls succeed, so flaky agents built on self-hosted open models may not be the model's fault.

At the hardware end, Nvidia says Kazakhstan is getting 45,000 Grace Blackwell GPUs, with Vera Rubin planned for next year. Nvidia also introduced Minerva, pitched as end-to-end observability for AI factories. For engineers who want to understand models below the API, Sebastian Raschka has released a from-scratch video course on building reasoning models.

Trust you inherit from CAs and vendors

A report found Indian government sites serving revoked eMudhra certificates. Codeminer42 discusses the eMudhra certificate incident to explain how CA trust propagates through root stores and what happens when that trust is questioned. A reachable endpoint can still present a certificate that should no longer be accepted, so revocation status needs to be checked directly.

Cydome says an n0n extortion incident affecting 38 maritime companies likely traces back to a single vendor, an attribution the firm presents as probable rather than confirmed. Wiz analyzed Gheist, an attempt to abuse stolen GitHub tokens at scale that ultimately stumbled, which illustrates how much damage a long-lived token can enable. Microsoft shipped out-of-band .NET security fixes for a bug in Google's Brotli implementation, so services that decode Brotli need prompt runtime patching. Unit 42's survey of current OAuth attacks is relevant to anyone maintaining delegated-access integrations.

Breaches can also have long aftermaths. Hackers claim to have stolen 590GB of PwC data, and the company has responded. UK Biobank is preparing to come back online months after someone tried to sell data on 500,000 people on Alibaba.

Staffing decisions and AI adoption

The Seattle Times reports that Microsoft is offering buyouts to roughly a quarter of its workforce. That figure describes eligibility, not departures. American Banker reports that American Express is promoting AI while planning a 5% downsizing, but the report does not establish how much of the cut is due to AI. Allen Hutchison makes the explicit argument for spending on tokens rather than hiring.

Other writers stress the conditions that AI tooling depends on. Ben Balter argues that documentation quality now determines AI output quality, because agents increasingly work from internal docs, and stale docs produce confidently wrong answers. Scrum.org recommends anchoring an AI rollout on an unavoidable project instead of demos. Replit's Monitoring Agent moves agents into watching production, so its alerting quality and trust boundaries should be compared with existing on-call tooling. When judging any of these programs, CIO.com's warning about a metric that flatters IT organizations and David Cummings on what to do when anecdotes contradict metrics may be useful checks.

Smaller infrastructure changes worth noting

AWS added 1:1 NAT to Transit Gateway and Cloud WAN, which addresses overlapping CIDR ranges after acquisitions or in partner links and may let some teams retire self-managed NAT appliances. Teams with voice, video or messaging on Microsoft's platform need a plan now that Azure Communication Services is retiring. TanStack introduced a new library, TanStack Redact. Stalwart shipped version 1.0 of its mail server, billing it as the fastest available, which is the project's own claim. For low-level work, a post shows how ESP32-S3 SIMD instructions can accelerate computer vision on a roughly $2 chip.