From the cryptocurrency community to AI: how does Hermes Agent challenge OpenClaw with security and self-evolution?

Author: Groove BlockBeats

Original title: Starting in the coin industry, why did Hermes Agent become OpenClaw's biggest challenger?

On February 25, 2026, Nous Research released Hermes Agent v0.1.0. After 42 days, on April 8, the project had iterated to v0.8.0, 8 major versions, merged hundreds of PRs, and 242 contributors. At the same time, OpenClaw, the most popular open source AI agent project on GitHub, had 346,000 stars, but it also accumulated 138 security flaws in 63 days.

The two growth curves are rising at the same time, but what is rising is completely different.

From its official launch on January 29th to March 3rd, OpenClaw only took 33 days to surpass React to become the most starred software project in GitHub history. According to OpenClaw Statistics, 34,168 stars poured in within 48 hours during the peak period, which is equivalent to 710 stars per hour. For reference, it took Kubernetes about three years to reach 100,000 stars.

However, according to Blink Security Blog tracking, security researchers were disclosing CVEs at an average rate of 2.2 per day in the same time window. A total of 138 cases were recorded in 63 days, including 7 severity levels (CVSS 9.0 and above) and 49 high risk levels, accounting for a total of 41%. The most damaging one is CVE-2026-25253, a CVSS 8.8 point zero-click remote code execution vulnerability. The attacker only needs to let the user visit a malicious web page to steal the authentication token through the WebSocket gateway and fully control the user's agent. According to Shodan scan data, more than 42,000 OpenCLAW instances were exposed on the public network in February, and 63% of them did not enable gateway authentication.

On February 14, Peter Steinberger, founder of OpenClaw, announced he joined OpenAI and handed over the project to the Open Source Foundation. Since then, the frequency of disclosure of security issues has further accelerated.

This is the background to Agent Hermes playing. Not a quiet racetrack, but a market where trust is collapsing. But understanding Hermes simply as an “OpenClaw replacement” misses more important information. The two projects have fundamental differences at the architectural level.

OpenClaw's skills are static Markdown files, handwritten by users and distributed through the ClawHub marketplace. According to a February audit by Snyk's security team, 1,467 of the 5,700 skills on Clawhub were identified as malicious, including credential theft, crypto mining, persistent backdoors, and prompt injection. Of these, 91% used a mix of prompt injection and traditional malware techniques. The maximum number of installs for a single malicious skill was over 340,000.

Agent Hermes took a very different path. Its skills are not written by users; they are generated by the Agent itself. After completing a complex task (which usually involves more than 5 tool calls), Hermes will refine the execution experience into a reusable skill document and store it as structured Markdown according to agentskills.io open standards. When similar tasks are encountered in the future, the Agent automatically invokes and optimizes these skills. Every 15 tasks automatically trigger a cycle of reflection to evaluate which skills are working and which need to be improved.

The memory system is also different from the bottom design. OpenClaw relies on three plain text files (soul.MD for personality, Memory.md for notes, and user.md for user portraits), and cross-session memory requires manual user configuration. Hermes has a built-in hierarchical persistence architecture: persistent note layer, FTS5 full text retrieval, Honcho user modeling, hot/cold storage separation, and supports 6 pluggable backends. Users don't need to manage anything manually; agents decide what to remember and what to forget.

The differences in security models are more straightforward. The default security configuration of OpenCLAW is described by security researchers as “weak”. Gateway authentication is disabled by default, and skill execution is not isolated in a sandbox. Hermes has built in prompt injection scanning, credential filtering, context scanning, and container hardening (read-only root filesystem + ability to discard) from day one. As of April 9, Hermes Agent had no public CVE records.

Simply put, OpenClaw is a “toolbox” and you tell it how to do it. Hermes is a “growing assistant” who learns how to do better by doing things.

The pace of iteration is also talking. In the 42 days from v0.1.0 to v0.8.0 of Hermes Agent, the v0.2.0 version merged 216 PRs, solved 119 issues, connected to 7 messaging platforms, and wrote 3,289 tests. According to GitHub data, 27,000 stars correspond to 242 contributors, and the contributor-to-star ratio is approximately 1:111, which means 1 out of every 111 followers is writing code, and the community participation intensity is much higher than OpenClaw.

More noteworthy is the team behind Hermes. Nous Research isn't a startup that just popped up out of nowhere. They began with the Discord community in 2022 and spent three years becoming one of the most influential players in the field of open source AI models. According to HuggingFace data, the Hermes series models have been downloaded more than 33 million times. From Hermes 1 in 2023 (LLaMA 13B fine-tuning, ranked first in multiple benchmarks) to Hermes 4 (70B parameters) in 2025, to Hermes Agent, the line is continuous: make a model first, then act as an Agent. Model ability is the foundation of Agent ability.

Their roots are in web3. CEO Jeffrey Quesnelle was previously the chief engineer of Eden Network, an Ethereum MEV infrastructure project. The January 2024 seed round was led by Distributed Global and OSS Capital, with Solana co-founder Raj Gokal personally participating. In April 2025, Paradigm, one of the biggest venture capital funds in the crypto sector, led a $50 million Series A, with a token valuation of $1 billion. Note that this is token valuation, not traditional equity valuation.

This means that Nous Research is native to Web3 from its governance structure to its technical architecture. Their Psyche network is built on the Solana blockchain and is a decentralized AI training infrastructure. Hermes 4.3, released in December 2025, is the first model to be trained entirely on the Psyche network, done with consumer-grade GPUs distributed around the world rather than relying on centralized data centers.

The web3 team is no exception when it comes to influencing the AI community. On March 31, an engineer named Chaofan Shou discovered an Anthropic Claude Code source leak. The absence of a .npmignore file resulted in 512,000 lines of TypeScript code being publicly published to npm. According to VentureBeat, the leaked mirror warehouse received 100,000 stars within 24 hours. Chaofan Shou's other identity is an engineer at Solayer Labs and co-founder of blockchain security company Fuzzland, a web3 security researcher who dropped out of UC Berkeley, and created one of the biggest 2026 code leaks in the AI world.

What Nous Research is doing is essentially similar: porting the methodologies trained by the web3 community (open source first, decentralized governance, community-driven iteration) to the AI Agent infrastructure layer. Hermes Agent's 42-day iteration rate of 8 major versions is partly a product of this methodology.

OpenCLAW's security crisis was a catalyst, but not the cause. The real variable is exactly how the AI Agent should be built. Whether to give users a toolbox to assemble themselves, or to build a system that can learn and evolve on their own. Nous Research answered the latter question in three years and 33 million model downloads, then turned the answer into a product in 42 days.

Twitter:https://twitter.com/BitpushNewsCN

Compare the TG exchange group:https://t.me/BitPushCommunity

Compare TG subscriptions:https://t.me/bitpush