Skip to content
NewCybercrime Report 2026 · Eight Chapters · A Post Is Not An Incident → Explore The ChaptersCybercrime Report 2026 → Explore

Odin Research / What is / AI tools and agents

What Is Hermes Agent? The Self-Improving AI Agent and Its Security Risks

An open-source agent that schedules its own work, writes its own skills and talks over chat apps: how it works, where its secrets end up, and how to keep it from leaking.

By
Odin by Brandefense
Updated
21 Nov 2026
Reading time
7 min read
Type
Autonomous AI agent

In short

Hermes Agent is an open-source, self-improving AI agent from Nous Research. It runs as a terminal app or a background service, uses dozens of tools such as a shell and a browser, schedules its own tasks and chats through Telegram, Discord and other apps. It is not malware, but exposed installs can leak API keys.

Key takeaways

What it is
An MIT-licensed AI agent from Nous Research that learns skills from experience, keeps memory and can run unattended with a built-in scheduler.
How it works
A model drives 40+ tools, per the project, including a shell and a browser. A messaging gateway lets users talk to it from chat apps, and its state lives in a hidden ~/.hermes folder.
Why it matters
An unattended agent is a service account that writes its own secrets to disk. Request dumps, snapshots and backups keep keys long after rotation.
What Odin Research saw
87 servers exposing Hermes Agent folders over open HTTP in Monograph 16. On 47 (54.0%), files showed schedules and task boards running without a person at the keyboard.
What to do first
Never serve a home folder over HTTP, deny /.hermes/ at the proxy, and delete dumps, snapshots and .env backups whenever you rotate a key.

What is Hermes Agent?

Hermes Agent is an open-source, self-improving AI agent made by Nous Research. It can run in a terminal or as a background service, use tools such as a shell and a browser, schedule its own jobs and chat through apps like Telegram and Discord. It is legitimate software, but careless installs expose API keys and bot tokens.

The project's GitHub README describes it as "the agent that grows with you". It creates skills from experience, refines them during use and searches its past conversations. It is published under the MIT license and had about 265,000 stars on GitHub on 21 November 2026.

Hermes Agent is one of several AI agents that developers now run on rented Linux servers, next to OpenClaw and Claude Code. Unlike a chat window, it keeps working when nobody is watching. That is its main appeal, and its main security challenge.

How Hermes Agent works

According to the README, Hermes Agent combines five parts:

  1. Model access. It works with Nous Portal, OpenRouter, OpenAI, custom endpoints and other providers. Users switch models with a command.
  2. Tools. The project lists more than 40 tools and supports MCP servers. In the installs Odin Research reviewed, Hermes registered 31 tools, including a shell, a headless browser, a scheduler and a message sender.
  3. Memory and skills. The agent curates its own memory and writes skills after complex tasks. The project says its skills follow an open agent skills standard.
  4. Scheduling and subagents. A built-in cron scheduler runs jobs and delivers results to messaging platforms. Isolated subagents work on tasks in parallel.
  5. Gateway and backends. A messaging gateway connects Telegram, Discord, Slack, WhatsApp, Signal and email. Commands can run locally or in Docker, SSH, Modal, Daytona or Vercel Sandbox backends.

On Linux, macOS and WSL2, Hermes keeps everything under ~/.hermes. The security documentation says secrets live in ~/.hermes/.env and configuration in ~/.hermes/config.yaml.

Hermes Agent security model

The project documents several built-in safeguards:

  • Command approval. A smart mode (the default) lets a helper model approve low-risk commands and escalate uncertain ones. A manual mode always asks. An off mode, and a "YOLO" switch, skip prompts. A hard blocklist still refuses commands such as wiping the root file system.
  • Fail-closed jobs. Cron, single-query and webhook sessions deny risky commands by default, because no human is there to answer.
  • Gateway access. With no allowlist set, the gateway denies all users. Unknown senders get a short pairing code that the owner must approve, and codes expire after one hour.
  • Secret handling. Sensitive variables are stripped from child processes, MCP error messages are redacted, and tools cannot write to protected paths such as .env.

The docs recommend explicit allowlists, a container or SSH backend for gateways, a non-root user and chmod 600 on the .env file. None of these controls stops an owner from publishing the whole folder with a quick file server, which is what Odin Research keeps finding.

Why Hermes Agent matters to defenders

Hermes Agent is not attack software. The risk comes from what an unattended agent leaves on disk and who can reach it.

  • A service account with a shell. An agent that runs commands on a timer and posts results to a chat bot holds real access. Treat it as a non-human identity, with scoped tokens and egress limits.
  • Secrets outlive rotation. Request dumps, upgrade snapshots and config backups keep full copies of keys. This is a textbook case of credentials in files.
  • Shadow deployments. Developers install agents on personal cloud servers outside IT's view. See shadow AI.

What Odin Research observed

Odin Research has studied exposed Hermes Agent installs in three reports. Every server below belonged to an owner who left a directory listing open. They are victims of their own configuration, and all addresses and secrets stay withheld.

  • 87 exposed servers. Monograph 16 reviewed 87 servers serving a Hermes folder over plain HTTP. On 47 (54.0%), files showed schedules, task boards and self-maintaining skill libraries running unattended. Port 8080 was the most common, on 38 servers.
  • What the folders held. Skills appeared on 68 hosts, logs on 57, Curator runs on 41, pre-upgrade snapshots on 6, cron jobs on 5 and task boards on 4. One Curator run made 90 tool calls, 12 of them shell commands, and cut its skills from 68 to 42 without a prompt.
  • Keys on disk. When a model provider returned a temporary error, Hermes saved the whole failed request, bearer header included. Odin Research verified 11 secret findings, including 3 TLS private keys, 2 Telegram bot tokens and keys for Anthropic, DeepSeek and Discord.
  • Noise from bundled files. Monograph 10 found Hermes on 75 of 112 exposed AI-agent servers, with 967 secret and session matches. Bundled skill packs ship example keys and local URLs, so scans must exclude /.hermes/skills/ before anyone opens an incident.
  • Keys shared across agents. On one server, the same Anthropic key sat in an OpenClaw profile and a Hermes snapshot. Monograph 09 showed two exposed Hermes gateways where a borrowed Claude Code login could land in the Hermes credential pool, an upstream feature that becomes a leak once the folder is public.

How to detect exposed Hermes Agent installs

  • Your attack surface. Search for HTTP 200 responses on /.hermes/ paths, and for directory listings that link to .hermes/.
  • Web logs. Requests for /.hermes/.env, profile .env or auth.json files, sessions/request_dump_*.json or state-snapshots suggest someone is harvesting keys.
  • Hosts. Alert on python -m http.server or similar listeners started inside home folders that hold .hermes, and on unattended Hermes gateway and Curator processes on servers that should not run agents.
  • Exclude skill mirrors. Skip /.hermes/skills/ and bundled documentation in secret scans, then compare any remaining key with the upstream package.

Monograph 16 publishes Suricata, Sentinel KQL, Sigma and YARA rules for Hermes daemon paths, retry dumps, task boards and cron stores. Monograph 10 adds a Suricata signature for requests under /.hermes/ that skips skill folders. All rules are in the Defense Kit.

How to run Hermes Agent safely

  1. Bind to loopback. Never start a file server in /root or /home bound to all interfaces. Deny /.hermes and other dotfile paths at every reverse proxy.
  2. Purge on rotation. When you rotate a model or bot token, also delete request dumps, pre-update snapshots and .env.bak files. See How to respond to leaked credentials.
  3. Keep approvals on. Use smart or manual approval, set gateway allowlists and run the gateway as a non-root user in a container.
  4. Register unattended jobs. Treat scheduled agent jobs and task boards as service accounts with owners, egress allowlists and short-lived tokens. See AI Agent Identity and Access and Monitoring and Logging for AI Agents.

Hermes Agent compared with other AI agents

Exposure figures come from Monograph 10, which counted secret and session matches across 112 exposed servers.

AgentHome folderServers (with user files)Matches
Hermes Agent/.hermes/75 (58)967
Claude Code/.claude/26 (25)988
OpenClaw/.openclaw/21 (19)932
Cursor IDE/.cursor/6 (3)184
OpenAI Codex CLI/.codex/16 (14)74

Hermes appeared on the most servers but with few matches per server, largely because of bundled skill files. Claude Code and OpenClaw showed fewer servers but more secrets each.

Frequently asked questions about Hermes Agent

What is Hermes Agent?

Hermes Agent is an open-source, self-improving AI agent from Nous Research. It uses tools such as a shell and a browser, keeps memory, writes its own skills, schedules jobs and chats through Telegram, Discord and other apps.

Is Hermes Agent malware?

No. Hermes Agent is legitimate MIT-licensed software. The risk Odin Research documents comes from owners exposing its ~/.hermes folder over HTTP, which can leak API keys and bot tokens.

Who made Hermes Agent?

Nous Research publishes Hermes Agent on GitHub as NousResearch/hermes-agent.

Where does Hermes Agent store API keys?

The documentation says secrets live in ~/.hermes/.env and configuration in ~/.hermes/config.yaml. Odin Research also found keys in request dumps, pre-update snapshots and .env backup files.

Is Hermes Agent safe to use?

It can be, if you keep command approval on, set gateway allowlists, run it as a non-root user in a container and never serve its folder to the internet. Delete dumps and snapshots when you rotate keys.

Can Hermes Agent run commands on its own?

Yes. Its cron scheduler and messaging gateway let it run tasks without a person present. The docs say unattended cron and webhook sessions deny risky commands by default when no human can approve them.

How do I find exposed Hermes Agent servers?

Check your external attack surface for HTTP 200 responses under /.hermes/ and for directory listings that link to .hermes. Odin Research publishes Suricata, Sigma, YARA and KQL rules for this in its Defense Kit.

Hermes Agent vs OpenClaw: what is the difference?

Both are open-source agents that developers run on servers. In Odin Research's data, Hermes appeared on more exposed servers with fewer secrets each, while OpenClaw showed fewer servers with more secrets per server.

Sources

  1. Nous Research, Hermes Agent repository and README, GitHub. github.com/NousResearch/hermes-agent (checked 21 November 2026)
  2. Hermes Agent documentation. hermes-agent.nousresearch.com/docs/ (checked 21 November 2026)
  3. Hermes Agent documentation, Security. hermes-agent.nousresearch.com/docs/user-guide/security (checked 21 November 2026)

Related tools and terms

All Odin Research explanations · Brandefense glossary

Odin Research

Get the rules, not just the definition

Every Odin Research report ships with verified indicators and detection rules you can load today.

Open the Defense Kit