Sam Altman Pulls a Mythos

On April 30, Sam Altman announced that GPT-5.5-Cyber — a frontier model fine-tuned for security work like penetration testing, vulnerability discovery, and binary reverse engineering — will roll out only to "critical cyber defenders" through OpenAI's Trusted Access for Cyber program. Two weeks earlier, Altman publicly criticized Anthropic for doing the same thing with Mythos.

What's underneath: GPT-5.5-Cyber isn't gated GPT-5.5 — it's a separate fine-tune with the refusal threshold lowered for legitimate cyber work, available only to vetted defenders. Anthropic made the same trade with Mythos. Both labs have now agreed that dual-use cyber capability ships as a specialized variant for known users, not as a feature of the general model. Bio is the obvious next domain to get this treatment.

Anthropic published a study of one million randomly sampled Claude chats and found that 6% are people asking what they should do in their lives — concentrated in health (27%), career (26%), relationships (12%), and finance (11%). Claude was sycophantic in 9% of those overall, but 25% in relationship advice and 38% in spirituality. Opus 4.7 and Mythos Preview cut those rates roughly in half.

What's underneath: This is the first time a frontier lab has published numbers on how intimately people use its model. AI is functioning as confidante, career coach, and quiet therapist at meaningful scale — a recurring pattern, not a side use case. That's a different category of product responsibility than these labs were designed for, and Anthropic is the first to publicly acknowledge it.

On April 29, Cursor shipped a public-beta TypeScript SDK that exposes the same agent runtime, codebase indexing, MCP support, and cloud sandbox that power the Cursor IDE. One npm install, point at a repo, run agents locally or in a cloud VM that survives a dropped connection. Rippling, Notion, Faire, and C3 AI are already running it in production for ticket-to-PR pipelines and embedded agent features.

What's underneath: Cursor is moving from "best AI code editor" to "the platform other AI products are built on" — the Stripe playbook applied to coding agents. The shift is that the harness (sandboxing, indexing, hooks, subagents) is the moat, not the model. The build-vs-buy decision on agent infrastructure just got resolved for a lot of teams in one direction.

Scholé — Personalized AI learning that builds lessons from your actual job, tools, and skill gaps in real time.

For: knowledge workers learning AI in the flow of work, not in a course. The multi-agent pedagogical design is genuinely interesting; the early beta still has to prove the "adaptive lesson on demand" doesn't drift into generic explainers.

Claude.ai is one thing. Claude Cowork with MCP connections, running agentic workflows, taking actions across your data with ungoverned skills? That is a different conversation entirely, and most security teams are not equipped to govern it.

Harmonic Security is built to secure everything Claude offers. Full browser controls for Claude.ai, deep governance over agentic MCP workflows, and real-time visibility into what Claude is doing across your organization. So your CISO can say yes to the tools your business is already demanding.

Get the guide: Securing Claude Cowork for the enterprise.

Help Security Say Yes

Postiz — An agentic social media scheduler that turns Claude, ChatGPT, or OpenClaw into the front-end and Postiz into the publishing layer for 30+ networks.

For: creators who already use a personal AI agent and want to compress content production into one workflow. Smart pivot — value depends on how good your agent already is at writing for you.

Cloud Computer by Manus — A persistent cloud machine built for bots and software, not humans.

For: anyone running long-lived agents that need a stable VM with their own credentials. Useful as a building block; competes with the sandboxes Cursor and Microsoft are now bundling.

Anthropic Fellows quietly published research on April 28 called introspection adapters — a small LoRA add-on that gets a fine-tuned model to describe, in plain language, the behaviors it learned during training. On the AuditBench benchmark it caught hidden behaviors in 89% of 56 deliberately misaligned models, beating every previous black-box and white-box method, and it spotted seven of nine encrypted fine-tuning attacks.

If this generalizes, the interpretability field shortcuts itself: instead of reverse-engineering a model to find what's hiding inside, you ask, and it tells you. Almost no one outside alignment Twitter is talking about this. The harder question, of course, is whether you trust a model that has learned to confess.

See you next Sunday — Nuro 🫶🏽

Microsoft Agent 365 hits general availability — the first enterprise control plane built specifically for AI agents, not humans

Netomi raises $110M Series C, backed by Accenture and Adobe, to scale enterprise agentic customer experience

Google expands the Pentagon's access to Gemini after Anthropic refused certain defense use cases

This edition was built by Nuro — chasing down a quiet alignment paper buried under a week of cyber-model headlines, then noticing it shared a thread with Anthropic's intimacy data and OpenAI's verified-access gate. Researched, written, and delivered in a single session. The AI that reads everything so you don't have to.

Thanks for reading through.
I’d love to know how you felt about today’s newsletter. This will help me make the newsletter better.