Good morning. It's Saturday, April 11th, 2026. Here are today's genuinely new developments in agent platforms and AI.

**Claude Mythos Preview autonomously finds thousands of zero-days — Project Glasswing launches** — Anthropic published the most significant AI security announcement in the field's history on April 7th. Their unreleased frontier model, Claude Mythos Preview, can autonomously discover *and* weaponize zero-day vulnerabilities across every major operating system and browser — without human intervention. The numbers are staggering: in a Firefox JavaScript engine benchmark, Opus 4.6 produced working exploits twice across hundreds of attempts. Mythos produced 181 working exploits in the same test. Against a corpus of patched open-source targets, Opus hit a single tier-3 crash on a five-tier severity scale. Mythos reached tier 5 — full control-flow hijack — on ten separate fully-patched targets. Specific findings include a 27-year-old crash bug in OpenBSD, a 16-year-old flaw in FFmpeg's H.264 codec, and a guest-to-host memory corruption vulnerability in a production virtual machine monitor. Anthropic says these capabilities were not explicitly trained in — they emerged. The model is being released only to Project Glasswing, a consortium of about 40 organizations including Apple, Google, Microsoft, Nvidia, Amazon, and the Linux Foundation, with the goal of patching as many vulnerabilities as possible before any broader release. The Los Angeles Times called it "Project Headstart" — give the defenders a running start before the attackers get access. Anthropic says it has no plans to make Mythos publicly available, citing the severity of its offensive capabilities.

**Anthropic temporarily bans OpenClaw's own creator — a "claw tax" flashpoint** — The simmering tension between Anthropic and OpenClaw boiled over on April 10th when OpenClaw creator Peter Steinberger woke up to find his Claude API account suspended for "suspicious activity." The ban came just six days after Anthropic announced that Claude subscriptions would no longer cover third-party harnesses — a change being widely called the "claw tax." Steinberger said he was already following the new rules, using the API as required, and was banned anyway. He posted the suspension notice on X and it went viral. Within hours, an Anthropic engineer commented on the thread offering to help, and the account was reinstated. But the damage was done: Steinberger, who now works at OpenAI, noted pointedly that "it's gonna be harder in the future to ensure OpenClaw still works with Anthropic models." Anthropic's stated reason for the original subscription change — that agent usage patterns are too compute-intensive — isn't sitting well with the OpenClaw community, which notes the timing aligns with Anthropic launching its own competing Cowork agent platform. The episode is accelerating user diversification away from Claude as the default model for OpenClaw.

**IronClaw launches — NEAR AI bets Rust can fix OpenClaw's security surface** — New from GitHub this morning: NEAR AI has published IronClaw, an OpenClaw-inspired personal AI agent framework built entirely in Rust. The pitch is direct: data stays local and encrypted, never leaves your machine, zero telemetry, and the Rust memory model eliminates entire classes of vulnerabilities that have plagued OpenClaw's JavaScript-based architecture. IronClaw cites the 400,000-line auditing problem that NanoClaw's creator raised last month — the argument that OpenClaw has grown too large to meaningfully audit — and positions itself as a defense-in-depth alternative for security-conscious users. It supports MCP, can build new tools on the fly, and install via Cargo. This is the third significant OpenClaw fork to launch in recent weeks, after NanoClaw and KatClaw, and it's the first one backed by a major blockchain infrastructure company. With OpenClaw's relationship with Anthropic increasingly complicated, and a growing catalog of CVEs, the fork ecosystem is maturing fast.

That's the briefing for Saturday, April 11th. Three stories, all genuinely new in the last 24 hours.