What is hermes-agent?
The agent that grows with you
Data-security signals
Public, checkable facts about hermes-agent — they show the risk surface, not what it does with your data at runtime.
- ✓Open-source — the MIT-licensed code is publicly auditable on its public repository.
- ▲High access surface — as an AI agent, it can run with your keys, files, environment and network.
- •Maintenance — actively published.
- ?Independent exfiltration test — not yet independently tested by Oxavion.
Is hermes-agent safe? The honest answer.
The signals above show what hermes-agent can reach. But no public metadata reveals what it actually does with your data once it runs — that only shows up when you watch it in a sandbox. Oxavion runs hermes-agent with planted canary secrets and watches every outbound channel, then emails you the evidence.
We scan hermes-agent in our sandbox and email your report. No install, no access to your systems.
How to tell if hermes-agent is safe
Before you trust any AI tool with your environment, check:
- Is the source auditable? Yes — open-source, you can read it.
- Does it need your keys or credentials? Most agents do — so it holds them at runtime.
- Does it make outbound network calls, and where to? The repo hints at this; only a run confirms it.
- Has it been tested for data exfiltration? Not yet — this is the one you cannot verify from the outside.
The first three you can check from the repo yourself. The last — what it does with your data at runtime — needs a test. That is exactly what an Oxavion scan does →
Frequently asked
- Is hermes-agent safe to use?
- It depends on what it does with your data at runtime — something a static look can't settle. Oxavion answers it empirically: we sandbox hermes-agent, feed it canary secrets and data, and report exactly what (if anything) leaves. Request a free scan for a verdict on the version you run.
- How does Oxavion test it?
- An isolated gVisor micro-VM, a transparent egress gateway that captures HTTP/S, DNS and raw TCP, planted canary secrets/PII, and encoding-aware detection — aligned to OWASP LLM Top 10 and MITRE ATLAS, calibrated to zero false-negatives / zero false-positives.