Key Shifts

  • Anthropic files for IPO + launches Project Glasswing: Confidential IPO filing follows a $65B raise at a ~$1T valuation. Simultaneously launches Project Glasswing, a consortium spanning Apple, AWS, Google, Microsoft, NVIDIA, CrowdStrike, and others. Claude Mythos Preview found thousands of zero-days — including a 27-year-old OpenBSD vulnerability and a 16-year-old FFmpeg bug. Samsung, SK Hynix, and SK Telecom are among the 15+ countries now receiving Mythos access for critical infrastructure 출처: TechCrunch · Anthropic
  • Enterprise AI ROI crisis surfaces: Uber blew its annual AI budget in four months and imposed a $1,500/month per-employee cap. COO: “it’s very hard to draw a line between AI usage and new consumer features.” On HN, “AI Doesn’t Have ROI” essay hits 56 points as skepticism spreads 출처: TechCrunch · Wheresyoured
  • Agent workplace race accelerates: Microsoft unveils Scout, an always-on autonomous agent integrated with M365 apps. OpenAI ships six job-specific Codex plug-ins (data analytics, sales, investment banking, creative production, product design, equity investing). Codex hits 5M+ weekly active users; knowledge workers are 20% of the base and growing 3x faster than developers 출처: Computerworld · TechCrunch
  • AI cybersecurity arms race: Claude Mythos autonomously chained multiple Linux kernel exploits for privilege escalation. OpenAI counters with GPT-5.5-Cyber. 404 Media reports Microsoft internal docs aim to “make people addicted” to Scout 출처: 404 Media

Startup / Product / Platform Radar

  • Anthropic: Opus 4.8 with “dynamic workflow” tool. Claude Mythos expanded to 15+ countries’ critical infrastructure. Confidential IPO filing
  • Microsoft: Scout (autonomous agent), MAI-Code-1-Flash and MAI-Thinking-1 new models, agent behavior control tools, text-based AI behavior testing tools
  • OpenAI: Six Codex job plug-ins + Sites (hosted web output) + Annotations. GPT-5.5-Cyber for security
  • GitHub Copilot: Switches to usage-based token billing, developer community erupts (“What a joke”) 출처: Ars Technica
  • DeepSeek-V4-Flash: Successfully running on AMD MI300X
  • Google: Fake call detection against AI deepfake impersonation scams
  • DuckDuckGo: ‘No-AI’ search engine traffic surges, accessibility enhanced 출처: TechCrunch
  • YC P26: Expanse (unlock wasted GPU capacity), Rudus (AI for concrete contractors)
  • Also: Skillhound (public SKILL.md access), vLLM session-aware agentic routing, RSS revival driven by AI agents

AI Future Signals

  • AI-native security auditing becomes infrastructure: That Mythos found a 27-year-old OpenBSD vulnerability implies vast undiscovered flaws remain in the software we all depend on. Within 1–2 years, continuous AI-driven security auditing of all critical infrastructure will be standard, not optional
  • End of the “free money” AI budget era: Uber is the canary. Copilot’s usage-pricing backlash signals that both enterprises and developers are realizing AI costs far exceed initial expectations. The real barrier to AI adoption won’t be technical capability — it’ll be economics
  • Always-on agents become the new OS layer: Scout’s “autopilot” concept — agents working in the background without being prompted — combined with Google Spark and OpenAI Codex, represents a battle for the agent OS layer. Whoever wins this layer owns the next decade’s platform
  • AI Cold War blocs forming: Project Glasswing (Anthropic + US Big Tech alliance) vs. China (DeepSeek). What began in cybersecurity is evolving into geopolitical alignment around AI standards, infrastructure access, and allied-nation lines

Realistic Opportunities / Experiments

  • AI cost optimization middleware: Every company with AI spend needs what Uber just scrambled to build — usage monitoring, model routing (expensive vs. cheap models), budget alerts, team-level quotas. Copilot billing backlash confirms market timing is right
  • Vertical agents with hard ROI metrics: Vague “AI boosts productivity” claims are dead. Build agents for narrow domains where cost savings or revenue lift is directly measurable — construction (Rudus), manufacturing, logistics. The differentiator is the number, not the demo
  • AI security audit SaaS for the mid-market: Mythos-level capability is out of reach, but there’s a gap below the enterprise/government tier. Mid-size companies can’t afford a Glasswing partnership but still need vulnerability scanning — a SaaS play with clear per-audit pricing

Uncertainties / Keep Watching

  • Will Anthropic’s IPO reprice the entire AI sector? $1T valuation may hold in private markets but public markets are less forgiving
  • Is Uber an outlier or the leading edge of an enterprise AI spending correction? If it’s the latter, Copilot, Cursor, Codex revenue models face a reckoning
  • Could Project Glasswing evolve into a de facto AI security alliance? If cybersecurity expands into AI standards and regulation, non-participating companies and nations risk exclusion
  • How durable is the ’no-AI’ backlash? DuckDuckGo’s surge, data center moratoria, Harvard commencement rhetoric — temporary reaction or a lasting consumer segment?