Skip to content
Source · Daily Brief

AI Daily Brief — 27 March 2026

Friday opens with a sharp cybersecurity-stock selloff as the markets react to Thursday’s Anthropic Mythos leak. CrowdStrike falls ~7%, Palo Alto Networks ~6%, Zscaler ~4.5%; the Global X Cybersecurity ETF (BUG) closes at its lowest since November 2023, more than 21% off year-to-date. Bloomberg, Benzinga and CoinDesk all lead with the fallout. Two Chinese coding labs ship — Zhipu’s GLM-5.1 and KwaiPilot’s KAT-Coder-Pro V2 — while Tom’s Hardware pegs NVIDIA Vera Rubin NVL72 rack pricing at $7-8.8M per rack and a Challenger report names AI as the single largest stated reason for US layoffs in March.

Top stories

  • Cybersecurity stocks plunge on Mythos. CrowdStrike -7%, Palo Alto -6%, Zscaler -4.5%, with Okta, SentinelOne and Fortinet each near -3%. The Global X Cybersecurity ETF (BUG) closes at its lowest since Nov 2023; Bitcoin and broader software baskets slip in sympathy as traders price in unknown-exploit risk and a potential reset of cybersecurity spend. via Bloomberg
  • Anthropic confirms Mythos. The company confirms to Fortune that the leaked Claude Mythos is real — “a step change” and “the most capable we’ve built to date.” Anthropic says training is finished and the model is in early-access piloting. The leak is traced to roughly 3,000 assets left in a publicly accessible CMS folder via human configuration error. via SiliconANGLE
  • Raymond James — “AI compresses cyber defences.” Adam Tindle publishes a Friday note: traditional defensive approaches based on known signatures, vulnerability databases and prior threat-intel telemetry “could be pressured as AI enables continuous discovery of unknown exploits.” The note flags compression of defensive advantages, higher attack complexity and inevitable shifts in security architecture and spending. via Benzinga
  • Z.ai — GLM-5.1 launches. Zhipu ships GLM-5.1: a 754B-parameter MoE (40B active), 200K context. Third-party coding eval puts it at 45.3 on Claude-Code — about 94.6% of Claude Opus 4.6’s 47.9, a 28% jump over GLM-5. API live the same day; open-source weights are slated for 7 April. GLM Coding Plan starts at a $3/mo promo. via Binance Square
  • KwaiPilot — KAT-Coder-Pro V2. Kuaishou’s lab ships an agentic-coding MoE with 72B active params, 256K context, output up to 80K tokens, ~109 tok/s — and hits 79.6% on SWE-Bench Verified. The architecture: a “Specialize-then-Unify” paradigm distilling SWE, WebCoding, Terminal, WebSearch and General experts into one model. via Puter

Money & infra

Tom’s Hardware reports NVIDIA Vera Rubin NVL72/NVL144 rack systems landing in the $7-8.8M range — with HBM4 + LPDDR5X memory alone at ~$2M (about 25% of total BOM) and Rubin GPUs at only ~$50K apiece, squeezing OEM/ODM margins hard. Arm + Meta + Google + NVIDIA unveil an “AGI CPU” — the first data-centre processor specifically designed for agentic AI workloads. Both Intel and AMD signal CPU price hikes of up to 15% on the AI-driven server-CPU shortage; Broadcom begins shipping Tomahawk 6 networking silicon for AI back-end fabrics. Separately, the New Mexico State Land Office blocks the natural-gas pipeline that would have powered the planned 2.45 GW Stargate Project Jupiter campus on the US-Mexico border. via Tom’s Hardware

Policy & labour

NY Governor Kathy Hochul signs amendments to the state’s RAISE Act, shifting the frontier-AI law toward a transparency- and reporting-based framework and stepping back from the earlier outright deployment restrictions — a New York counter-positioning against the White House’s 20 March federal-preemption push. Challenger, Gray & Christmas‘s March job-cuts report names AI as the single largest stated reason for March cuts: 15,341 AI-attributed layoffs in the month, ~25% of the total — a sharp escalation from under 8% in 2025. OpenAI‘s ad pilot reportedly crosses $100M ARR in less than two months, with 600+ advertisers. via Challenger

Quiet corners

No major shipping from Google DeepMind, Meta AI, Microsoft AI or Thinking Machines Lab. DeepSeek stays silent through a fourth rumoured March window — V4 Preview will not land until 24 April. Import AI‘s slot is empty: issue 450 published Monday, 451 will land Monday 30 March. The Mistral Voxtral TTS analysis wave — open-weights TTS preferred over ElevenLabs Flash v2.5 in 68.4% of blind zero-shot voice-cloning tests — keeps trending into Friday coverage.

By the numbers

  • Cybersecurity ETF BUG: lowest close since Nov 2023, >-21% YTD
  • GLM-5.1: 754B / 40B active · 200K context · 45.3 Claude-Code (94.6% of Opus 4.6)
  • KAT-Coder-Pro V2: 72B active · 256K context · 79.6% SWE-Bench Verified · ~109 tok/s
  • Vera Rubin NVL72 rack: $7-8.8M · HBM4 + LPDDR5X ~$2M (25% of BOM) · Rubin GPUs ~$50K each
  • March AI-attributed US layoffs: 15,341 (~25% of all March cuts) — up from <8% in 2025
  • OpenAI ad pilot: $100M+ ARR in <2 months · 600+ advertisers

Compiled by AI Feed’s editor from publicly reported announcements on 27 March 2026.