# State of the Claw

Peter Steinberger, OpenAI | AI Engineer Europe 2026 | 44:12

Source: https://www.youtube.com/watch?v=zgNvts_2TUE
Channel: AI Engineer (https://www.youtube.com/@aiDotEngineer). Summarised by AIE Talks.
Page: https://aietalks.com/talks/state-of-the-claw
Published: 2026-04-17
Tags: agents, coding-agents, open-source, security, team-adoption

## TL;DR
- OpenClaw's rapid growth has created a serious maintenance burden, especially around security reports and contributor turnover.
- Peter Steinberger wants OpenClaw to remain independent, open, and compatible with both hosted and local models through the OpenClaw Foundation.
- Building useful agents still depends on human judgment, taste, system design, and the ability to say no to attractive but disconnected ideas.

## Summary
Peter Steinberger gives an update on OpenClaw after five months of rapid growth. He describes a project with tens of thousands of commits, nearly 2,000 contributors, and a security workload that has become difficult for volunteers to handle. He explains why many alarming advisories do not describe practical attacks against the recommended setup, while also acknowledging real risks around agents that access private data, untrusted content, and communication tools. The OpenClaw Foundation is intended to bring in full-time maintainers and keep the project independent from any single company, including OpenAI. In the Q&A, Steinberger discusses local models, his multi-agent coding workflow, the importance of taste and personality, agents embedded throughout the home, prompt injection, and experimental features such as dreaming. He is direct about the limits of automation. Agents can write much of the code, but people still need to choose the direction and understand the whole system.

## Key ideas
### OpenClaw's growth has created an unusually large project and a management problem
[01:03](https://www.youtube.com/watch?v=zgNvts_2TUE&t=63s)
Steinberger says OpenClaw is five months old and has grown in a straight line rather than the usual hockey-stick curve. He describes it as the fastest-growing project in GitHub's history, with around 30,000 commits, nearly 2,000 contributors, and almost 30,000 pull requests. The project is still gaining velocity. That growth has also left him with two jobs, working at OpenAI and running the OpenClaw Foundation. He calls the foundation a company on hard mode because volunteers cannot be directed like employees. One of his goals is improving the bus factor, including finding more people who can handle commits. Contributors from Nvidia, Microsoft, Red Hat, Tencent, ByteDance, and other companies are helping.

### Security reports need careful assessment because severity scores often hide the practical risk
[03:47](https://www.youtube.com/watch?v=zgNvts_2TUE&t=227s)
Steinberger says OpenClaw received 1,142 security advisories, about 16.6 per day, with 99 marked critical. The project published around 469 and closed 60 percent of them. He compares the volume with the Linux kernel and curl, while saying that the larger problem is how reports are interpreted. In his rule of thumb, the more loudly an issue is described as critical, the more likely it is to be slop. He gives an example involving a CVSS 10 issue where a restricted read-only setup could be turned into a writable one, but says almost nobody uses that permission model. Other problems are real, including supply-chain issues and malware distributed through fake packages. Nvidia's security work also found several ways to escape a sandbox.

### OpenClaw's recommended setup limits several attacks, but powerful agents still carry real risk
[10:03](https://www.youtube.com/watch?v=zgNvts_2TUE&t=603s)
Steinberger argues that some coverage treats theoretical configurations as if they were the normal installation. OpenClaw recommends keeping a personal agent out of group chats, enabling sandboxing for team agents, and giving a team agent only the data that the team should know. The default gateway token is local-only, or private-network-only when remote access is needed. He says agents remain exposed when they can access private data, consume untrusted content, and communicate with other systems. That combination is a general risk for agentic software, not a problem unique to OpenClaw. His position is conditional: more capable agents can do more, but users must understand what those capabilities allow.

### The OpenClaw Foundation is meant to support independence and paid maintenance
[16:22](https://www.youtube.com/watch?v=zgNvts_2TUE&t=982s)
Steinberger rejects the idea that OpenAI bought OpenClaw. He says OpenAI understands that the project needs to remain open and work with any model, including local models. He is building the OpenClaw Foundation as an independent structure, inspired by Ghostty, and says it is close to completion. A delay with the American banking system is holding it up. The foundation should make it possible to hire full-time people who can keep up with the project's pace, improve its quality, and free Steinberger to work on new features. He is deliberately adding people from Nvidia, Microsoft, Telegram, Salesforce, Tencent, ByteDance, and other companies, because bringing in only OpenAI employees would make the project look controlled by OpenAI.

### Local models matter because users should keep control of their data
[22:28](https://www.youtube.com/watch?v=zgNvts_2TUE&t=1348s)
Steinberger says control over personal data was one of the reasons he built OpenClaw. He is wary of systems that connect a company's model directly to a user's Gmail while the email is hosted elsewhere. His preferred setup keeps most data under the user's control and sends only what is needed to a top-tier model. He also describes a fallback hierarchy of models. OpenClaw's agent can interact with websites directly, which lets consumers get around some of the integration silos that make Gmail access difficult for startups. He supports broad model compatibility, but warns that small local models may not have enough defensive training for use with email or a web browser. OpenClaw warns users when they select one.

### Steinberger uses several concurrent coding agents, but he still develops iteratively
[24:57](https://www.youtube.com/watch?v=zgNvts_2TUE&t=1497s)
Steinberger says he sometimes ran almost ten coding sessions at once, especially when using Codex with GPT-5.1, and now more often uses five or six windows because the loops are faster. He treats this as a temporary workaround until token generation becomes faster. He rejects a fully automated dark-factory approach in which the developer specifies everything at the beginning and never reviews the result. Software ideas change while they are being built, and experimentation reveals shortcuts and new directions. He builds in steps, runs the software, changes his prompts, and evaluates what happens. Pipelines can automate parts of the work, but merging every pull request would let the AI pull the product in directions that do not fit its purpose.

### Taste includes avoiding obvious AI patterns and caring about small product details
[28:28](https://www.youtube.com/watch?v=zgNvts_2TUE&t=1708s)
For Steinberger, the lowest level of taste is recognizing when something smells like AI. He points to writing style, personality, and user interfaces that repeat familiar patterns such as a colored border or a purple gradient. The judgment is often immediate, like recognizing AI-written slop even when it is hard to explain exactly why. Since more software work can be automated, he thinks developers have more time to refine small details. He gives OpenClaw's occasional teasing messages as an example of a detail that adds character and would not emerge from a high-level prompt alone. He also describes adapting the agent's writing for WhatsApp, where the default model felt too wordy and unlike the way his friends text.

### Agents need human direction because system design and saying no remain difficult
[40:24](https://www.youtube.com/watch?v=zgNvts_2TUE&t=2424s)
Steinberger says engineers should keep developing taste and system design skills. An agent can produce good or bad code depending on whether the human asks the right questions and defines sensible boundaries. He also says saying no is becoming more important because every idea is only a prompt away. One idea is manageable, but many ideas can combine into a system that no longer fits together. The agent often works from an incomplete view of a codebase, perhaps relying on an outdated agent.md file and the files immediately in front of it. Humans need to provide project-level hints, ask how a change interacts with other parts, and keep the whole system maintainable.

## Notable quotes
- "The more they scream how critical they are, the more likely it's slop." (05:02)
- "The more powerful you make it, the more it can do for you, but the more you also have to understand what it does." (14:33)
- "The bottleneck is still syncing and having taste." (28:05)
- "I think a skill that is becoming more and more important is saying no." (42:31)

## Tools & references mentioned
- OpenClaw
- OpenClaw Foundation
- OpenAI
- Nvidia
- Microsoft
- Red Hat
- Tencent
- ByteDance
- Salesforce
- Telegram
- Slack
- Codex
- NemoClaw
- Linux kernel
- curl
- Agents of Chaos
- Simon Willison
- Andre Karpathy
- Maran Dure

## Who should watch
- You maintain a fast-growing open-source project and need a candid account of security triage, volunteers, contributors, and foundation-building.
- You are deciding whether agents should use local models, private data, browsers, or communication tools and want Steinberger's practical risk boundaries.
- You are using coding agents and need a concrete explanation of where iterative judgment, product taste, and system design still matter.

## Related talks

- [Claws Out: Securing and Building with OpenClaw](https://aietalks.com/talks/claws-out-securing-and-building-with-openclaw) (Nick Taylor, Pomerium, 17:12)
- [AIE Europe Keynotes & OpenClaw](https://aietalks.com/talks/aie-europe-keynotes-openclaw) (Phil Hawksworth & Lia McBride, AI Engineer & Malte Ubl, Vercel & Raia Hadsell, Google DeepMind & Ryan Lopopolo & Peter Steinberger, OpenAI & swyx, AI Engineer & Vincent Koc, Comet ML & Radek Sienkiewicz, VelvetShark & Sally Ann O'Malley, Red Hat & Nick Taylor, Pomerium & Onur Solmaz, OpenClaw & Merve Noyan, Hugging Face & Fryderyk Wiatrowski, Viktor & Gergely Orosz, The Pragmatic Engineer & Kitze, Sizzy & Matt Pocock, AI Hero & Sunil Pai, Cloudflare, 9:11:01)
- [Dark Factory: OpenClaw Ships Faster Than You Can Read the Diff](https://aietalks.com/talks/dark-factory-openclaw-ships-faster-than-you-can-read-the-diff) (Vincent Koc, OpenClaw, 16:44)
- [Develop at Idea Velocity](https://aietalks.com/talks/develop-at-idea-velocity) (Jeffrey Lee-Chan, Snapchat, 15:28)
- [A Piece of Pi: Embedding the OpenClaw Coding Agent in Your Product](https://aietalks.com/talks/a-piece-of-pi-embedding-the-openclaw-coding-agent-in-your-product) (Matthias Luebken, Tavon, 20:42)