The Two Faces of OpenClaw

With Peter Steinberger's TED talk and AIE talk released on the same day, OpenClaw appeared to the public as an inspiring success story, while looking entirely different to engineers.

On the public stage, the focus was on how OpenClaw came to be and its achievements, but on the technical front, the story was far rougher. Security issues were severe, with 60x more security reports and an estimated at least 20% of skill contributions being malicious, along with major scalability challenges in maintaining the fastest-growing open source project in history.

This contrast is the key point. One is "the story of building something that seemed impossible," while the other is the cost of the operations, security, and quality control that actually underpinned that growth.

The same thread continues more broadly afterward in the AI Twitter Recap.

• Anthropic Claude Opus 4.7 and Claude Design were released, sparking discussion not only about model performance but also about expansion into a design/prototyping tool.
• As Computer use and subagents rapidly became practical, reactions followed emphasizing that real-world agent harnesses and evaluation methods are becoming important.
• On the research side, work such as Cognitive Companion, WebXSkill, and Autogenesis—addressing agents' self-improvement, accumulation of web skills, and closing performance gaps—drew attention.

Overall, OpenClaw is presented as both a success story and a case showing how quickly operational, security, and evaluation problems are growing in the age of AI agents.