In partnership with
The Boy Who Cried Step Function
For two years, “too dangerous to release” has been marketing theater. GPT-4 shipped. The world kept spinning. We learned to roll our eyes.
This week, the wolf might actually be at the door.
Mythos: The Model Behind the Curtain
On April 7, Anthropic announced Claude Mythos Preview under Project Glasswing — no public API, restricted to a consortium of ~12 named partners (AWS, Apple, Microsoft, Google, CrowdStrike, Palo Alto Networks, Cisco, NVIDIA, JPMorgan Chase, Broadcom, the Linux Foundation) plus additional critical-infrastructure orgs. Stated scope: defensive cybersecurity only.
Why the lockdown? Mythos has autonomously discovered thousands of zero-days across every major OS and browser — production code that survived decades of human review and fuzzing:
27-year-old OpenBSD bug — remote crash via signed integer overflow in TCP/SACK handling.
17-year-old FreeBSD NFS RCE (CVE-2026-4747). Unauthenticated root from the internet. ROP chain built across multiple packets, zero human steering.
16-year-old FFmpeg H.264 flaw that survived ~5M automated fuzzing iterations.
Composed Linux kernel escalations — KASLR bypass + heap manipulation + root, as a sequence.
181 successful Firefox 147 exploits, up from 2 with Opus 4.6.
Fast browsing. Faster thinking.
Your browser gets you to a page. Norton Neo gets you to the answer. The first safe AI-native browser built by Norton moves with you from idea to action without slowing you down. Magic Box understands your intent before you finish typing. AI that works inside your flow, not beside it. No prompting. No copy-pasting. No switching apps.
Built-in AI, instantly and for free. Privacy handled by Norton. Built-in VPN and ad blocking protect you by default. No configuration. No extra apps. Nothing to think about.
Fast. Safe. Intelligent. That's Neo.
Benchmarks back the jump: CyberGym 83.1% (Opus 4.6: 66.6%). SWE-bench Verified 93.9% (80.8%). SWE-bench Pro 77.8% (53.4%). Category change, not version bump.
99%+ of findings remain unpatched. Anthropic is using SHA-3 commitments to prove future disclosure without exposing details today. Commitment: $100M in compute credits to Glasswing partners, $4M to open-source security orgs. Polymarket: ~4% odds of public release by April 30.
GPT-4 could write a better phishing email. Mythos autonomously probes production systems, composes multi-stage exploits, and lands root. The framing is still marketing. The capability is no longer theater.
This Week’s Public Shipments
Claude Opus 4.7 + AI design tool + Claude Code 2.0 — all this week (per The Information, April 14).
Design tool: natural language → websites, decks, landing pages, prototypes. Figma partnership for editable handoff. Integrated into Word and PowerPoint. Pre-launch reaction: Figma −6%, Wix −4.7%, Adobe −2.7%, GoDaddy −3%.
Claude Code 2.0: multi-session sidebar, drag-drop panes, integrated terminal, in-app editor. Plus Routines — scheduled agents, API endpoints, GitHub webhooks running on Anthropic’s infra. Cron for Claude. Pro/Max/Team/Enterprise.
The Accidental Preview
The March 31 npm leak: 513K lines of unobfuscated TypeScript across 1,906 files, 44 feature flags, an always-on background agent KAIROS, and autoDream (memory consolidation during idle). Internal model codenames Capybara and Tengu exposed via “Undercover Mode.” Anthropic called it a “plain developer error.”
Security footnote: a concurrent supply-chain attack shipped a malicious axios (1.14.1 / 0.30.4) with an RAT to anyone who updated Claude Code via npm between 00:21–03:29 UTC on March 31. If you did — audit. Native installer is now the recommended path.
And the other shoe: GPT-5.5 “Spud”. Pretraining finished ~March 24. Altman: “very strong model.” Brockman: “two years of research… big model feel, not incremental.” Polymarket: 78% by April 30, 95%+ by June 30. Naming (5.5 vs 6) still unconfirmed; OpenAI is reportedly waiting on the size of the performance delta.
The Read
Three weeks. Multiple releases. The step-function is arriving as 0.1 bumps — and most of what’s moving now never shows up on a model card. Mythos is hidden. KAIROS is implied. Routines is infrastructure. If you’re only watching headline benchmarks, you’re watching the wrong tape.
Subscribe to AI Insider to read the rest.
Become an AI Insider to get access to this post and other subscriber-only content.
Join AI InsiderAI Insider members get:
- ✅ Full access to 100% of all content.
- ✅ Exclusive DEMOs, reports, and other premium content.
- ✅ Ad-free experience.
