AI News Roundup — 2026-04-11 (Enterprise + Product)
Anthropic's Mythos found thousands of zero-day vulnerabilities, but smaller, cheaper models matched much of its capability — exposing the myth of AI moats via model scale alone. Meanwhile, OpenAI faced infrastructure security lapses (Axios compromise), and the push for enterprise AI infrastructure continued (Cloudflare's EmDash for agents, Berkeley's work on trustworthy benchmarks).
Top Stories
- AI Cybersecurity Gets Real (But Model Size Isn't the Moat) — Anthropic's Mythos found 27-year-old and 16-year-old zero-days across major OSes; AISLE tested those same vulnerabilities on 3.6B parameter models and recovered the same analysis for 90% of cases. Source
- Cirrus Labs Joins OpenAI — The AI security research firm that builds autonomous vulnerability-detection systems is being acquired to strengthen OpenAI's red team capabilities. Source
- OpenAI Suffers Software Supply Chain Hit — Axios HTTP library compromise (March 31) affected ChatGPT Desktop, Codex, and Atlas signing certificates; OpenAI updating security certs. Source
- Berkeley Exposes "Jagged Frontier" in AI Benchmarks — Small and large models show wildly inconsistent capability across security reasoning tasks; no single model dominates all domains, challenging claims of capability scaling. Source
Shipping & Platform
- Cloudflare Builds WordPress for AI Agents (EmDash) — New content platform designed as a foundation for AI agents to publish autonomously; highlighting WordPress's stagnation in the AI era. Source
- Gmail E2EE Now on Mobile — Enterprise end-to-end encryption available on Android and iOS apps for organizations with the feature enabled. Source
- Microsoft Removing Copilot Buttons from Windows — Unnecessary Copilot integrations disappearing from Notepad, Snipping Tool, Photos, and Widgets after user pushback. Source
Policy & Governance
- 20-Year-Old Arrested for Alleged Attack on Sam Altman's Home — Suspect threw Molotov cocktail at CEO's house and allegedly made threats at OpenAI offices. Source
- AI-Generated Iranian Lego Videos Go Viral — Explosive Media's AI-driven geopolitical meme campaign is sparking White House-level response; highlighting AI's role in information warfare. Source
- France Switching Government to Linux — DINUM announced exit from Windows to reduce dependence on non-European tech; broader sovereignty play affecting enterprise AI infrastructure choices. Source
One Take
Mythos's landmark zero-day discoveries matter—but not because of model size. AISLE's findings expose a critical blind spot in AI capability discussions: the real moat is the system (expertise, tooling, integration), not the weights. For enterprises betting billions on frontier model licensing, this is humbling. Meanwhile, infrastructure moves (EmDash, France's Linux shift, Google's E2EE mobile rollout) signal the enterprise market is consolidating around platform stickiness and sovereignty. Action item: If you're evaluating AI security tools or building agent infrastructure, demand evidence that capability translates across model families—don't assume the priciest model is the safest.