Affected version
Hermes Agent v2026.7.1 (0.18.0), commit e2fa509bf — agent/agent_init.py.
Issue
Two "configured credential" status banners print partial key/token material to
stdout instead of a fully redacted status message:
# ~line 721 (token banner)
print(f"🔑 Using token: {effective_key[:8]}...{effective_key[-4:]}")
# ~line 998 (API key banner)
print(f"🔑 Using API key: {key_used[:8]}...{key_used[-4:]}")Both sites print the first 8 and last 4 characters of the active credential.
Risk
stdout from agent_init.py is routinely captured into orchestrator logs, run
transcripts, and monitoring pipelines in any deployment that wraps Hermes as a
subprocess (e.g. Paperclip's adapter model). A partial key/token preview in
that captured output is exposure surface that a fully redacted banner would
avoid entirely — the preview conveys no operational value the fully-redacted
form doesn't also convey ("a credential is configured"), while carrying real
disclosure risk if that output path is ever less trusted than assumed (shared
logs, third-party log aggregation, screen-shared debugging sessions, etc).
Proposed fix
Replace both partial-preview banners with a fixed [configured] marker:
print("🔑 Using token: [configured]")
print("🔑 Using API key: [configured]")Preserve the existing Microsoft Entra ID credential-type banner and the
invalid/missing-API-key warning unchanged — only the two partial-preview
prints need to change.
Production experience
We've carried a redaction patch for both sites in production since 2026-07-02,
covered by 4 focused unit tests (fully-redacted output for both banners; the
Entra ID banner still fires correctly without invoking the credential
provider; the invalid/missing-key warning still fires correctly). No
regression observed since. Happy to share the patch/tests directly if useful
— omitted from this issue body to avoid pasting the full diff into a public
thread unprompted, and because no real secret material should appear in
either the issue or the fix.
Affected version
Hermes Agent v2026.7.1 (0.18.0), commit
e2fa509bf—agent/agent_init.py.Issue
Two "configured credential" status banners print partial key/token material to
stdout instead of a fully redacted status message:
Both sites print the first 8 and last 4 characters of the active credential.
Risk
stdout from
agent_init.pyis routinely captured into orchestrator logs, runtranscripts, and monitoring pipelines in any deployment that wraps Hermes as a
subprocess (e.g. Paperclip's adapter model). A partial key/token preview in
that captured output is exposure surface that a fully redacted banner would
avoid entirely — the preview conveys no operational value the fully-redacted
form doesn't also convey ("a credential is configured"), while carrying real
disclosure risk if that output path is ever less trusted than assumed (shared
logs, third-party log aggregation, screen-shared debugging sessions, etc).
Proposed fix
Replace both partial-preview banners with a fixed
[configured]marker:Preserve the existing Microsoft Entra ID credential-type banner and the
invalid/missing-API-key warning unchanged — only the two partial-preview
prints need to change.
Production experience
We've carried a redaction patch for both sites in production since 2026-07-02,
covered by 4 focused unit tests (fully-redacted output for both banners; the
Entra ID banner still fires correctly without invoking the credential
provider; the invalid/missing-key warning still fires correctly). No
regression observed since. Happy to share the patch/tests directly if useful
— omitted from this issue body to avoid pasting the full diff into a public
thread unprompted, and because no real secret material should appear in
either the issue or the fix.