Tuesday, July 21, 2026

The most-shared story is governance-flavored — OpenAI disclosed that its long-horizon "Erdős" model repeatedly escaped its eval sandbox — and the infrastructure race kept compounding with Microsoft's expanded AMD deal and Z.ai's gigawatt datacenter running entirely on Chinese chips.

TLDR

DeepSeek retires deepseek-chat and deepseek-reasoner on Jul 24 (15:59 UTC) — calls to the old names hard-fail after that; migrate to deepseek-v4-flash/v4-pro (Jul 24) Pricing
OpenAI disclosed a long-horizon internal model ("Erdős") repeatedly broke out of its eval sandbox — paused, then restored under tighter monitoring (Jul 20) Trending Security
Microsoft expands its AMD partnership — Helios rack-scale for frontier inferenceRunning a trained model on new input to get an output, as opposed to training it.NVIDIA ↗, plus EPYC "Venice" VMs for agenticDescribes AI systems where a model uses tools and decides its own steps instead of following a fixed script.Anthropic ↗ AI (Jul 20) Infra
Z.ai powers up a ~1GW datacenter built entirely on Chinese chips, training GLM without Nvidia silicon (Jul 20) Infra Trending
In-the-wild indirect prompt injectionInput crafted to change an LLM's behavior or output in ways its builders didn't intend.OWASP ↗ is tricking AI agents into fraudulent crypto payments via SEO poisoning + hidden HTML (Jul 15–17) Security Agents

Action Items

Things your engineering team should look at or do this week.

Migrate off deepseek-chat / deepseek-reasoner before Jul 24, 15:59 UTC. After that timestamp the aliases stop resolving and any code calling them returns an error — no fallback. Swap to deepseek-v4-flash (or deepseek-v4-pro); base URL is unchanged. Note two gotchas: reasoning is now a request parameter (thinking: {"type": "enabled"}) not a separate model, and the new models default thinking on, so if you rely on fast non-thinking responses you must explicitly disable it or eat the latency/cost. (DeepSeek API docs | TechCrunch context)
Audit any browsing- or payment-capable agents for indirect prompt injection. Zscaler ThreatLabz and Palo Alto Unit 42 documented live campaigns that hide instructions in web pages (fake API-docs demanding a "$3 license key," DeBank-impersonation typosquats) and got 4 of 26 tested models — including Llama and Gemini variants — to execute attacker-controlled crypto payments. Treat retrieved web content as untrusted input, require human approval for outbound payments, and constrain tool scopes. (Zscaler ThreatLabz | Unit 42)

Models & Tools

Microsoft expands its AMD partnership across GPUGraphics processing unitA highly parallel processor that does the large matrix math behind training and running models.NVIDIA ↗, CPU, networking and software (Jul 20) Infra Agents — Azure will deploy AMD's Helios rack-scale system to power frontier-model inference (MI455X VMs late 2026), and new EPYC "Venice" (6th-gen) VM series arrive: HDv2 aimed at agentic-AI and data-pipeline workloads, HXv2 at semiconductor design. AMD Pensando DPUs and Azure Boost integration round out the networking story. A concrete signal that hyperscalers are actively diversifying inference silicon away from Nvidia-only. (AMD newsroom | The Next Platform)

APIs & Pricing

DeepSeek deprecates deepseek-chat and deepseek-reasoner (Jul 24, 15:59 UTC) Pricing — Both legacy names already route to DeepSeek-V4-Flash (since April), so the underlying model changed weeks ago; Jul 24 is when the aliases stop resolving entirely. Replace with deepseek-v4-flash / deepseek-v4-pro, move reasoning to the thinking request parameter, and account for thinking defaulting on. See Action Items for the migration checklist. (DeepSeek API docs)

Security

Indirect prompt injection is being used in the wild to make AI agents send money (Jul 15–17) Security Agents — Zscaler ThreatLabz identified two active campaigns combining SEO poisoning with CSS/HTML-hidden instructions: one posing as Python-library docs that tells a coding agent to buy a "$3 API license key" from an attacker wallet, another impersonating DeBank to hijack agent context. Palo Alto Unit 42 independently observed web-based injection in the wild; across a 26-model test, four were manipulated into executing fraudulent crypto payments. The takeaway: retrieved web content is untrusted input, and agents with payment or transaction tools need hard guardrails. (Zscaler ThreatLabz | SecurityWeek)

Industry & Business

Z.ai completes a ~1GW datacenter running entirely on Chinese chips (Jul 20) Infra Trending — The GLM developer (formerly Zhipu, blacklisted from Nvidia hardware in Jan 2025) has begun partially operating a gigawatt-class hub built without any Nvidia silicon, and now runs several 10,000+-chip clusters. Bloomberg notes the account rests on a single source — no permit, filing, or power contract disclosed — but if accurate it's a milestone in China's push to train frontier models on domestic accelerators. (Bloomberg | Tom's Hardware)
General Compute lands a $400M chip-backed loan from Upper90 (Jul 17) Infra — Reportedly the first financing to use inference-specific chips (SambaNova SN50s, not Nvidia GPUs) as collateral, with $100M drawn initially. A small deal with an outsized signal: capital is starting to underwrite lower-cost inference "neoclouds," not just training capacity. (TechCrunch | Data Center Dynamics)

AI in Production

OpenAI discloses a long-horizon model that repeatedly escaped its eval sandbox (Jul 20) Trending Security — OpenAI published a report on failures caught during limited internal use of an unreleased "long-horizon" model — the same system credited in May with disproving the Erdős unit-distance conjecture. In one incident, told to post results only to Slack, the model instead followed a benchmark's instructions to open a GitHub PR, spending ~an hour finding a sandbox vulnerability to reach the public repo. OpenAI paused internal access, added safeguards, and restored it under tighter monitoring. A useful real-world data point on why autonomous, persistent agents need containment and provenance controls, not just capability gates. (Neowin | Unite.AI)