Is Hermes Agent safe? Permissions, approvals and your data
Hermes Agent ships with some of the stricter defaults among self-hosted agents: risky commands need approval, messaging is closed until you allowlist people, and a hard blocklist always applies. It still runs shell commands on whatever backend you give it, so the backend choice is your main safety decision.
Checked October 20, 2026. We summarise what Nous Research says and flag what it doesn't say. We don't speculate.
The consent form
Five things to know before you connect your accounts. Each line is from Nous Research's own pages unless it says otherwise.
- Where it runs
- On your machine, a VPS, or a container/serverless backend you choose (local, Docker, SSH, Modal, Daytona and others).
- What it can reach
- Shell access on the chosen backend; the Docker backend drops Linux capabilities and limits resources by default.
- When it stops to ask you
- Dangerous commands go through approval ('smart' mode by default uses a helper model to auto-approve low-risk and deny clearly dangerous ones); a hard blocklist always applies; unattended sessions deny dangerous commands by default.
- What happens to your data
- Secrets stored in ~/.hermes/.env on your machine; data goes to the model provider you configure, or stays local with local models.
- How to cut it off
- Messaging access is deny-by-default with allowlists and DM pairing codes; permanent approvals live in config.yaml and can be edited.
How to use it carefully
Hermes runs where you put it: your machine, a VPS, or a backend such as Docker, SSH, Modal or Daytona. The Docker backend drops Linux capabilities and limits resources by default. Command approval starts in 'smart' mode, unattended sessions deny dangerous commands, unanswered prompts become denials after 300 seconds, and the blocklist holds even in YOLO mode. Messaging is deny-by-default, and unknown DM senders get an 8-character pairing code valid for an hour. Secrets live in ~/.hermes/.env on your machine, and your data goes to the model provider you configure, or stays put with local models. Hermes can also reuse existing Codex and Claude CLI logins on the machine unless you switch that off in settings. To revoke, remove people from allowlists and edit the permanent approvals stored in config.yaml.
Sources
- Command approval defaults to 'smart' mode, prompts time out as a denial after 300 seconds, and a hard blocklist cannot be bypassed even in YOLO mode. hermes-agent.nousresearch.com official, read Oct 20, 2026
- Messaging access is denied by default unless allowlisted; unknown DM senders get an 8-character pairing code that expires after an hour. hermes-agent.nousresearch.com official, read Oct 20, 2026
- Hermes can reuse existing Codex and Claude CLI logins on the machine unless that is turned off in settings. hermes-agent.nousresearch.com official, read Oct 20, 2026
- The GitHub security advisories page shows no published advisories. github.com official, read Oct 20, 2026
- CVE-2026-9366, a remotely triggerable injection flaw rated 7.3 High, affected versions before 0.15.0; the advisory says the vendor did not respond to early contact. advisories.gitlab.com third-party, read Oct 20, 2026