Skip to content
NewCybercrime Report 2026 · Eight Chapters · A Post Is Not An Incident → Explore The ChaptersCybercrime Report 2026 → Explore
All short reads
Short readAI agent exposure4 min read

Unattended AI Agents Are Leaving Their Keys on Open Servers

By
Odin by Brandefense
Published
19 November 2026
Reading time
4 min read

Key takeaways

  • Odin found 87 servers exposing Hermes Agent folders over open HTTP.
  • 47 of them (54.0%) show unattended automation, not interactive chats.
  • Failed model requests are saved to disk with the bearer header inside.
  • One open file server also exposed TLS private keys for two web panels.
  • Bind file servers to loopback, move certificate keys, and purge dumps on key rotation.

In a two-week observation window ending 18 November 2026, Odin by Brandefense reviewed 87 servers that expose a Hermes Agent folder over plain HTTP. Hermes is an open-source AI agent that can run as a background service. On 47 of these servers (54.0%), the exposed files show schedules, task boards and self-maintaining skill libraries running without a person at the keyboard. Several also hold API keys, bot tokens and, in one case, TLS private keys. For defenders, this is a new kind of exposure: a service account that writes its own secrets to disk.

All 87 servers belong to users who left a directory listing open. They are victims of their own configuration, not attacker infrastructure. We withhold every address, domain and secret.

What does an unattended agent leave behind?

Hermes keeps its state in a hidden ~/.hermes folder. On a busy install, that folder holds far more than chat history.

Figure 1: Self-directed Hermes daemons coordinate four unattended subsystems
Figure 1: Self-directed Hermes daemons coordinate four unattended subsystems
SubsystemWhat it doesHosts (of 87)
Skills libraryPlaybooks the agent follows, plus bundled docs68
LogsGateway and agent logs57
CuratorThe agent reviews, merges and archives its own skills41
Language serverCode navigation across the machine7
State snapshotsCopies of .env and config taken before each upgrade6
Cron schedulerTimed scripts and model briefings sent to Telegram5
Kanban boardsTask cards handed to isolated worker agents4

One server in Germany exposes all eight subsystems in 14,914 files. Two old WebUI backups and 26 Curator entries show it has been upgraded and used for a long time.

On a Korean server, one Curator run took 591 seconds. The agent made 90 tool calls, including 12 shell commands, and cut its skills from 68 to 42. Nobody prompted it. This is ordinary housekeeping, but it shows how much an unattended agent does on its own.

The agent's toolset explains why this matters. Hermes registers 31 tools, including a shell, a headless browser, a scheduler and a message sender. An agent with those tools can run commands on a timer and post the results to a chat bot. Some installs also hand coding work to Claude Code or OpenAI Codex, or read and send email through a terminal mail client.

How do keys end up on disk?

Three Hermes features keep secrets long after an owner thinks they are gone.

  • Retry dumps. When the model provider returns a temporary error, Hermes saves the whole failed request. That file holds the bearer header, the agent's persona and memory, and its 31 tool definitions. One dump was 256,668 bytes.
  • Upgrade snapshots. Before each update, Hermes copies .env and config.yaml into a snapshot folder. Older Anthropic, OpenRouter, Telegram and Discord keys stay there.
  • Config backups. Edits leave .env.bak and config.yaml.bak files behind, each a full plaintext copy.

On one US server, a user asked the agent to report its fallback model without showing any key. The upstream model then failed, and the gateway saved the full request, bearer header included, to disk.

We verified 11 secret findings across these hosts. They include 3 TLS private keys, 2 Telegram bot tokens, and one key each for Anthropic, DeepSeek and Discord.

The lesson is simple. Rotating a key is not enough. The dumps, snapshots and backups must be deleted too.

Why does one open folder expose the whole server?

Most of these servers run a simple HTTP server on a high port: 38 of the 87 use 8080. When that server starts in /root, it serves everything next to .hermes as well.

Figure 4: Flagship servers pair autonomous Hermes stacks with co-hosted TLS keys and PM2 apps
Figure 4: Flagship servers pair autonomous Hermes stacks with co-hosted TLS keys and PM2 apps

On the US server, a Python script started in /root to show a dashboard. It exposed 7,643 Hermes files, 2,222 scheduler files, an OpenAI Codex install with 688 files, and backups. It also exposed 277 certificate files from acme.sh. Among them were the ECDSA private keys for two web-panel domains. Anyone with those keys could pose as the panels until the certificates are revoked.

The files cannot tell us whether a person or the agent started that file server. Either way, the result was the same.

Is every exposed Hermes folder a problem?

No. The skills folder is the most common path, at 499 of 1,100 directory entries. Much of it is bundled documentation. Our analytical model puts 25 hosts (28.7%) in a static group with only skills or a minimal config.

Logs (57 hosts) and Curator runs (41 hosts) are better signs of a working agent. The hosts that need action show task databases, schedules, retry dumps or private keys. Some hosts also carry red-team skill templates. These fit authorized testing or research benches as well as misuse, and nothing here shows an attack on a third party.

What to do now

  • Bind to loopback. Stop any http.server or similar listener in /root or /home that is not bound to 127.0.0.1, and deny /.hermes and /.acme.sh at the web edge.
  • Move certificate keys. Keep acme.sh keys outside home folders, readable only by their owner. Revoke any certificate whose key was exposed.
  • Purge on rotation. When you rotate a model or bot token, delete request dumps, pre-update snapshots and .env backups too.
  • Govern unattended agents. Register scheduled agent jobs and task boards as service accounts, with egress allowlists and short-lived tokens.
  • Deploy the rules. The full report's Detection rules tab has Suricata, Sentinel KQL, Sigma and YARA rules that skip bundled skill folders.

The full report covers the subsystems, file sizes, rule logic and hunting patterns in detail.

What is…?

Short explanations of the tools and families named in this study.

All explanations

Frequently asked questions

What does an unattended agent leave behind?

Hermes keeps its state in a hidden ~/.hermes folder. On a busy install, that folder holds far more than chat history. One server in Germany exposes all eight subsystems in 14,914 files. Two old WebUI backups and 26 Curator entries show it has been upgraded and used for a long time.

How do keys end up on disk?

Three Hermes features keep secrets long after an owner thinks they are gone. On one US server, a user asked the agent to report its fallback model without showing any key. The upstream model then failed, and the gateway saved the full request, bearer header included, to disk.

Why does one open folder expose the whole server?

Most of these servers run a simple HTTP server on a high port: 38 of the 87 use 8080. When that server starts in /root, it serves everything next to .hermes as well.

Is every exposed Hermes folder a problem?

No. The skills folder is the most common path, at 499 of 1,100 directory entries. Much of it is bundled documentation. Our analytical model puts 25 hosts (28.7%) in a static group with only skills or a minimal config.

Read the full report · 20 min read

Unattended AI Agents Exposed: Hermes Kanban Boards, Cron Jobs and Leaked Keys

In a two-week observation window ending 18 November 2026, Odin by Brandefense reviewed 87 servers exposing Hermes Agent folders over open HTTP. On 47 of them (54.0%), the files show ...

Full report and detection rules