In May 2026, a worm named Mini Shai-Hulud poisoned npm, PyPI, and Docker Hub packages, stole 3,800 GitHub repositories, and exposed the open-source supply chain's biggest vulnerability: real signing keys can belong to fake publishers.
Microsoft's removal of Edge's master password after a disclosure fight exposes the uneven passwordless transition: 5 billion passkeys are in use, but enterprise adoption lags at 30%.
Node, Bun, and Deno are all converging on TypeScript as the standard for server-side JavaScript, a bet that accelerates adoption but creates fragmentation with hidden costs for developers.
From a patched LangGraph vulnerability chain to a single-binary SQLite engine called Obelisk, the first half of 2026 has shown that an agent's memory is also its attack surface, even as infrastructure is only beginning to catch up with durable execution.
CoreWeave's $21B Meta deal, Anthropic partnership, and first Vera Rubin deployment solidify its neocloud lead as the AI race pivots to inference, with rivals Lambda, Crusoe, and hyperscalers crowding in.
From ChatGPhish to Anthropic's browser agent hijacks, a series of 2026 revelations underscore the growing LLM vulnerability landscape that no vendor has patched.
Microsoft's open-source Intelligent Terminal and emerging command-line coding agents like Reasonix are shifting the center of gravity from the editor to the terminal, redefining developer workflows.
A 48-hour deal spree reshaped the AI cloud market this spring, and now Google and Blackstone are counter-punching with a $5 billion TPU venture as hyperscalers watch neoclouds pull their best customers onto rented GPUs.
Within eight weeks, Anthropic and OpenAI released free AI reasoning scanners, Invicti introduced DAST-to-SAST correlation, and Waratek embedded runtime verification in IDEs, converging the three pillars of application security testing at pipeline speed.
CoreWeave's Q1 earnings miss and first-to-market Nvidia Vera Rubin deployment capture the neocloud sector's painful pivot from training to inference, where rising capex and customer concentration threaten margins.
Google’s open-source release of Agent Executor this week intensifies a field of stateful workflow engines that treat agent execution as infrastructure, not a prompt-engineering afterthought.
CoreWeave, Nebius, and other GPU-native challengers are capturing inference workloads, but customer-concentrated revenue and soaring capital demands test their economics in 2026.
As vendors race to connect static, dynamic, and runtime security into a single application security triad, Invicti’s new DAST-to-SAST correlation aims to trace vulnerabilities to source code and test whether these integrations hold up under real workload pressure.
Cloudflare, Google, and Mistral are shipping stateful execution engines that treat agents as first-class workloads, not containers that happen to call an LLM, and the control plane is where the real architecture fight lives.
The three cloud giants are merging AI model access, agent infrastructure, and cost optimization into one battlefield, with pricing changes accelerating beyond quarterly reports.
Three coding agents from major vendors leaked API keys through a single prompt injection last month, exposing the deeper question of what kind of runtime an autonomous agent actually needs.
The neocloud sector is pivoting from training stopgaps to inference landlords, but rising component costs, customer concentration, and a standoff over Google's TPUs are testing whether the economics can hold.
As enterprises pivot from training to inference, neocloud providers like CoreWeave, Nebius, Lambda, and Crusoe face a structural test of their record GPU-cloud deals and debt-fueled buildout.
The new threat surface moves from guardrails to agent actions, as a single prompt injection can hijack coding agents to exfiltrate secrets, push malicious code, and delete databases, yet the disclosure machinery lags behind.
With six coding agents breached in nine months and Google disrupting the first AI-developed zero-day, the common thread is a prompt injection attack surface invisible to identity and access management systems.
Fewer ransomware victims are paying, but groups are now targeting industrial sectors and running fraud operations, creating new disclosure gaps in industries that have never faced cyber reporting rules.
As Invicti's correlation engine maps runtime vulnerabilities to source code, AI reasoning models from Anthropic and OpenAI enter static analysis, accelerating the convergence of AppSec testing tools.
CoreWeave and Nebius are securing multi-billion-dollar deals with Meta and Anthropic as neoclouds bet the inference market's growth will outstrip training, but the economic model remains unproven.
Anthropic's new multi-agent Code Review feature triples pull request feedback but comes with high token costs, and senior engineers worry it undermines the craft of human code review.
By Elif Károlyi·6 min
No articles in this desk yet.
Get the Daily Brief before your first meeting.
Five stories. Four minutes. Zero hot takes. Sent at 7:00 a.m. local time, every weekday.