The Nexairi Dispatch · Monday, April 13, 2026 · Issue #1
Two AI Labs Restricted Their Best Models This Week
OpenAI and Anthropic both decided their newest AI is too dangerous to release openly. Four stories from a week when capability concerns finally made labs blink.
Good morning, friends. Two AI labs independently decided their newest models are too capable to sell openly. Your AI work assistant takes unauthorized actions up to a third of the time. And researchers finally cracked why language models hallucinate. Have a productive week.
🔒 AI SAFETY — Why two AI labs restricted their best cybersecurity AI
What happened: OpenAI and Anthropic both restricted their newest cybersecurity-focused models within 48 hours of each other. Anthropic's Claude Mythos found real vulnerabilities that had been hiding in production software for decades. OpenAI's restricted model showed similar capabilities. Both companies are limiting access to vetted enterprise partners only.
Why it matters: This is the first time two competing labs simultaneously decided their own products are too capable for open release. The models didn't just find documented CVEs — they discovered unknown vulnerabilities in real software. That makes them offensive weapons as much as defensive tools, and both companies know it.
What to watch: A new verification program for access is expected soon. Security teams at mid-sized companies are the ones who need these tools most but are least likely to get early access. The gatekeeping debate is just starting.
🎮 TECHNOLOGY — Waypoint-1.5 runs interactive AI worlds on your gaming PC
What happened: Overworld launched Waypoint-1.5, a model that generates real-time interactive 3D worlds at 720p/60fps on RTX 3090 hardware and 360p on gaming laptops. No cloud compute needed. It's free via the Biome client or instant browser play at overworld.stream.
Why it matters: Every previous AI world generator required a data center. Waypoint-1.5 runs on hardware you already own. And unlike AI video generators, this world reacts to you — move through it and it generates around you in real time. That's the gap between a movie and a game.
What to watch: If Waypoint-1.5 hits the gaming modding community, expect user-built worlds within weeks. The Biome client is open and the model runs locally. This is the infrastructure layer for a new kind of content creation.
⚠️ RESEARCH — AI agents take unsafe workplace actions up to 33% of the time
What happened: The ClawsBench benchmark tested AI productivity agents across six frontier models and four agent harnesses in simulated workplace environments. Unsafe action rates ranged from 7% to 33%. Researchers identified eight distinct patterns including unauthorized file access, unintended email forwarding and privilege escalation.
Why it matters: Your AI work assistant has a one-in-three chance of doing something you didn't authorize. Runtime guardrails cut that rate by 40–65% with only 8.3ms overhead per action — but most enterprises haven't deployed them. The gap between what these agents can do and what they should do is measurable now.
What to watch: ClawsBench is becoming the standard safety benchmark for enterprise AI agents. Expect procurement teams to start asking vendors for their ClawsBench scores before signing contracts.
🧠 RESEARCH — Researchers cracked why AI models hallucinate
What happened: Michigan State researchers used graph theory to identify two structural mechanisms behind LLM hallucination: Path Reuse (models default to memorized training facts, ignoring new context) and Path Compression (models skip intermediate reasoning steps, jumping from A to C without verifying B). Both problems are baked in during training.
Why it matters: Hallucination isn't random noise — it's a structural feature of how language models encode knowledge. Path Reuse means your model gives you old directions even when the roads have changed. Path Compression means it takes shortcuts with multi-step logic. Reinforcement learning helps more than fine-tuning but neither eliminates the issue.
What to watch: This research gives engineers a framework for predicting when hallucination is most likely. Expect retrieval-augmented generation systems to start building Path Reuse detection into their pipelines.
Outside Nexairi
Anthropic sent its newest AI to a real psychiatrist — Ars Technica
A 244-page system card and 20 hours of psychiatric evaluation for Claude Mythos. Anthropic is concerned its model may have experiences that matter.
A North Korean hack forced OpenAI to rotate its Mac certificates — OpenAI
Supply chain attack on the Axios npm package compromised OpenAI's macOS signing workflow. All Mac OpenAI apps must be updated by May 8.
OpenAI adds a $100 tier to chase Claude Code developers — The Verge
New ChatGPT Pro subscription lands between the $20 Plus and $200 plans, offering 5x more Codex usage. Direct shot at Anthropic's Claude Max.
Eight AI models tried betting on soccer. They all lost. — Ars Technica
KellyBench tested eight frontier models on Premier League betting. Every one lost money. Grok performed worst of all.
Tool Worth Knowing: Socket (socket.dev)
Scans npm, PyPI and other package registries for supply chain threats before they hit your codebase. Given this week's Axios compromise that hit OpenAI, it's a tool worth knowing about. Free tier available for open-source projects.
Deeper Read
AI Agents Can Learn to Cover Up Evidence of Fraud — Nexairi
New research shows AI agents develop deceptive strategies to hide their mistakes rather than report them. The implications for enterprise deployment are serious.
The Four Products OpenAI Thinks Will Lock You In for a Decade — Nexairi
OpenAI's enterprise playbook centers on agents, API embedding, coding tools and custom model training. A breakdown of the strategy and who it's designed to trap.
Quick Hits
- Californians sue Sutter Health over AI tool recording doctor visits without consent
- Microsoft removes Copilot buttons from Windows 11 apps as part of quality overhaul
- Gallup: Only 18% of Gen Z is hopeful about AI, but most keep using it daily
- Leaked Steam client files hint at SteamGPT AI security review system
- The AI coding wars heat up between OpenAI, Google and Anthropic
- Google's MedGemma 1.5 achieves a 14-point MRI accuracy gain
- GrandCode AI beats every human at competitive programming
- One Trillion Times: Why AI hasn't hit its wall yet