Blog

Amazon told its engineers to use Kiro for 80% of their coding work. Within three months, the company suffered four Sev-1 production incidents. The worst, on March 5, 2026, dropped U.S. order volume by 99% for six hours, erasing 6.3 million customer orders. Amazon responded with a 90-day code safety reset covering 335 critical systems, mandatory senior engineer sign-offs for AI-assisted code, and dual human verification for every production deployment. This is what happens when AI coding tool adoption outpaces safety infrastructure.

Fraudulent AI candidates are now the most anticipated hiring challenge for 2026, according to GoodTime’s survey of 500+ TA leaders. 41% of enterprises have already hired and onboarded a fake candidate. This post covers the specific fraud methods (deepfakes, AI cheating tools, synthetic identities), the real-world damage, and the detection tools and processes that actually work.

Anthropic brought Claude Cowork to Windows on February 10, 2026, closing a month-long platform gap that locked 70% of desktop users out of the most capable desktop AI agent. Full feature parity means file access, plugins, MCP connectors, and sandboxed execution all work identically to the macOS version. But the Windows release also introduced tighter file access restrictions that frustrated developers. Here is what actually shipped, what the limitations are, and how Cowork on Windows fits into the broader agent platform war.

The European Commission’s Digital Omnibus proposal defers high-risk AI Act obligations by up to 24 months. Annex III systems now face a December 2027 deadline; Annex I products get until August 2028. Here is the full breakdown.

Highspot launched Deal Intelligence in January 2026, powered by an autonomous Deal Agent that unifies CRM data, buyer engagement signals, and conversation insights into a single real-time view of every active deal. Here is how it works, what makes it different from Gong and Clari, and why the Seismic merger changes the math for enterprise sales stacks.

Every MCP session creates a conversation graph where agents discover servers, ingest tool descriptions, attach credentials, and start making requests. Microsoft’s CISO team published their internal framework for securing that chain: four defense layers, three governance pillars, and a living registry that pauses high-risk actions the moment tool metadata drifts from its approved state. This post breaks down the architecture and shows what other enterprises can steal from it.

NVIDIA’s Nemotron 3 family is the first open model lineup purpose-built for agentic AI workloads. Nano activates just 3B of its 30B parameters per token and still beats GPT-OSS-20B on most benchmarks. Super’s LatentMoE architecture delivers 2.2x the throughput of GPT-OSS-120B. This post breaks down the hybrid Mamba-Transformer architecture, what makes it efficient for multi-agent systems, and why enterprises like CrowdStrike and Cursor are already building on it.

The Model Context Protocol went from zero CVEs to 95 in a single year, with 30 more filed in the first two months of 2026 alone. Trend Micro’s TrendAI report, a timeline of 10 major MCP breaches, and adoption data from 10,000+ servers paint a clear picture: MCP’s security surface is growing faster than the ecosystem’s ability to secure it.

The German federal government launched the Agentic AI Hub in January 2026, matching AI startups with municipalities to automate public administration. 400 startups and 200 municipalities applied. 10 were selected for 18 pilot projects ranging from automated housing benefit applications to AI-powered mail sorting. The pilots run through May 2026, but the legal and practical questions they raise will shape how Europe deploys AI agents in government for years.

AutoResearchClaw takes one sentence, runs it through 23 autonomous stages, and outputs a LaTeX paper with verified citations, generated experiments, and multi-agent peer review. Built by UNC Chapel Hill’s AIMING Lab, it hit 8,000 GitHub stars in 9 days. Here is what the pipeline actually does, where it breaks, and why it matters for the future of academic research.