2026-05-25

1. Project Glasswing Initial Update — 1 month, 10,000+ vulnerabilities, fewer than 100 patches

• Anthropic released the first official status report on Project Glasswing on 2026-05-25 (source)
• Claude Mythos Preview, working with ~50 partners, found 10,000+ high-risk/critical vulnerabilities (including 1,000+ zero-days)
• Notable findings: a 27-year-old OpenBSD vulnerability and a 16-year-old FFmpeg vulnerability — both missed by existing automated tools and manual review
• Remediation Gap: 10,000+ vulnerabilities found vs. fewer than 100 patches deployed — AI detection speed far outpaces the industry's response capacity
• Why it matters: A structural crisis in which AI raises attack speed faster than defense speed. A growing database of "unpatched vulnerabilities" preemptively highlights what happens if Mythos-level capabilities are exposed to malicious actors.
• → Claude Mythos Preview, Anthropic

2. Google Deep Research Max launch — built on Gemini 3.1 Pro, MCP + private data integration

• Announced 2026-04-22 (first captured in this ingest) (source)
• A two-tier autonomous research agent built on Gemini 3.1 Pro: Deep Research (fast, conversational) vs. Deep Research Max (extended reasoning, asynchronous background)
• MCP support + private data: searches custom data sources such as internal documents and financial DBs together with the public web in a single run
• Performance: DeepSearchQA 93.3% (up from 66.1% in Dec 2025), HLE 54.6% (up from 46.4%)
• Announced alongside: Gemini Enterprise Agent Platform (an evolution of Vertex AI) — organization-level agent governance
• Why it matters: The MCP ecosystem is expanding beyond simple tool connections toward integration with private enterprise data. The research-agent competition among Anthropic (Stainless acquisition), OpenAI (existing Deep Research), and Google (Deep Research Max) is shifting to the "private data access" layer.
• → Deep Research Max, Google DeepMind

3. Gemini Robotics ER 1.6 — gauge reading 23%→93%, crossing the industrial autonomy threshold

• Announced 2026-04-15 (first captured in this ingest) (source)
• Enhanced Embodied Reasoning for physical AI — in particular 93% accuracy reading instrument panels (analog gauges) (a 4× improvement over ER 1.5's 23%)
• Co-developed with the Boston Dynamics Spot robot — demonstrated in real industrial scenarios (energy, chemical plants)
• Agentic vision pipeline: an observe → reason → verify loop
• Released to developers in the Gemini API and Google AI Studio
• Why it matters: Reading gauges and pressure meters is a core inspection task at energy, manufacturing, and water-treatment facilities. 90%+ accuracy is the usual threshold for practical deployment of supervised automation. DeepMind is converging with the "skill-library-based physical agents" direction Jim Fan (NVIDIA) laid out at CaP-X.
• → Gemini Robotics ER 1.6, Embodied Agents