State of the Claw — Peter Steinberger

Peter Steinberger, creator of OpenClaw and newly joined member of OpenAI, delivers a five-month state-of-the-project keynote at AI Engineer Europe 2026, followed by an audience Q&A moderated by Swyx. OpenClaw is described as the fastest-growing project in GitHub history — not a hockey-stick curve but a "stripper pole" straight line — with approximately 30,000 stars, approaching 2,000 contributors, 30,000 commits, and nearly 30,000 pull requests, with no slowdown in velocity.

Steipete dedicates a significant portion of the talk to addressing the security narrative around OpenClaw. The project has received 1,142 security advisories to date — roughly 16.6 per day — with 99 classified as critical, 469 published, and 60% closed. For comparison, the Linux kernel receives 8–9 advisories per day and curl has accumulated 600 total. He argues that the volume is inflated by AI-generated slop reports and security-industry incentives (researchers earn "credits" by discovering issues in high-profile projects). A concrete example: Nvidia's NemoClaw sandbox was broken out of in five different ways within 30 minutes using Codex Security and an unreleased model with stronger cyber capabilities. He details a CVSS-10 rated vulnerability — involving a read/write permission escalation in an unshipped iPhone sync feature — and explains why it is not a practical threat in normal deployments, criticizing how CVSS scoring ignores real-world deployment context.

Steipete pushes back hard on academic and media fearmongering, citing the "Agents of Chaos" paper that described OpenClaw's architecture in detail but omitted the security documentation page. He articulates the "legal trifecta" risk inherent to all powerful agentic systems: access to personal data + access to untrusted content + ability to communicate externally. This is not unique to OpenClaw; it is a property of any capable agent.

On governance, Steinberger explains his dual-role situation: full-time at OpenAI while running the OpenClaw Foundation, which he compares to "running a company on hard mode" with volunteers you cannot direct. The foundation is inspired by the Ghost CMS model and is nearly finalized, blocked only by the slowness of the American banking system for non-Americans. The goal is to hire full-time maintainers, reduce bus factor, and ensure OpenClaw remains model-agnostic and independent of any single company. He has brought contributors from Nvidia, Microsoft, Red Hat, Tencent, ByteDance, Telegram, and Salesforce, deliberately avoiding over-reliance on OpenAI engineers to prevent the appearance of a corporate takeover.

In the Q&A, Steinberger describes his parallel-agent coding workflow: running 5–6 Claude Code sessions simultaneously (down from nearly 10 during slower periods). He explains why full dark-factory automation fails — good software requires iterative discovery, not upfront waterfall specification. He defines "taste" as the ability to smell AI-generated slop in both writing and UI ("purple gradient" era), and as the attention to delightful details like startup messages that roast users.

On the future of agents, he envisions ubiquitous home agents accessible from any room, aware of the user's location, able to project information on the nearest display. He flags prompt injection as improving significantly at the frontier model level but still problematic with small local models that lack safety training. He teases work on "dreaming" — a sleep-cycle-inspired memory consolidation mechanism for agents, analogous to how humans convert short-term to long-term memory. Architecturally, OpenClaw has been refactored from spaghetti to a fully plugin-based system, enabling memory, wiki, dreaming, and other modules to be swapped independently — described as "Linux for agents."

On skills for engineers in the AI age, Steinberger highlights taste, system design (to avoid being "swiped into a corner" by localized AI solutions), and — crucially — the ability to say no. With any idea just a prompt away, the bottleneck is synthesis: understanding the whole system and steering agents toward architecturally coherent outcomes rather than accumulated local hacks.