Hermes Agent Hardening Guide
Harden Hermes Agent configuration authority, execution boundaries, credentials, integrations, tool access and private verification.
Overview
Nous Research Hermes Agent CLI/gateway, managed scope, dangerous-command approvals, MCP/plugins/skills, API/dashboard exposure and local state. These agents can expose source code, credentials and authenticated services to model-directed actions and third-party resources. Prioritize the authority available to the entire process, then restrict loading and individual tools.
Intended Audience
Security engineers, platform teams, IT administrators and responders governing developer or automation agents.
How to Use This Guide
- L1 (Crawl): Establish identity, conservative configuration, scoped credentials and private verification.
- L2 (Walk): Add whole-process isolation and externally enforced network restrictions for untrusted input.
- L3 (Run): Require independent evidence and change review for regulated deployments; this guide does not assert certification.
Scope
Nous Research Hermes Agent CLI/gateway, managed scope, dangerous-command approvals, MCP/plugins/skills, API/dashboard exposure and local state. The Hermes model family, hosted-business contractual controls and generic infrastructure hardening are outside scope. General OS hardening remains with your existing baseline. Isolation references here describe the agent-specific trust boundary, not a new host deployment recipe.
Research and validation boundary
Source and release checks were performed on 2026-11-15; the candidate is restricted to 0.23.1. See the source-first candidate and surface ledger for exact upstream commits, discovery gaps and evidence limits. This is documentation research plus offline hostile-input/native-schema/mock-CLI testing. No live agent/tenant was hardened, no device settings changed, and no live-validation status earned.
The Code Pack generates candidates and non-executing launch plans, compares explicit files and provides opt-in native reads. It does not install policy, provide a sandbox or compute complete effective state. All default operations are local/read-only or stdout-only. Native diagnostic execution is separate and explicit. Do not evaluate launch JSON as shell code.
Table of Contents
- Identity and Configuration Authority
- Execution and Network Boundaries
- Credentials and Integrations
- Tool and Data Restrictions
- Monitoring and Verification
- Compliance Quick Reference
- Appendices
1. Identity and Configuration Authority
1.1 Identify Nous Research Hermes Agent and pin its release
Profile Level: L1 (Crawl)
| Framework | Controls |
|---|---|
| NIST SP 800-53 Rev. 5 | SI-2, CM-8 |
Description
This guide covers Nous Research hermes-agent, not the Hermes model family or an unrelated package named Hermes. The candidate is tied to v0.23.1 and its release documentation; verify the deployed gateway/agent as well as the CLI.
Source: Official configuration/security reference, checked 2026-11-15. Recommendations and test procedures below are HTH’s interpretation of that interface and its limits.
Rationale
Why This Matters:
- Look-alike packages and stale examples can introduce a different trust model.
- Release-sensitive controls can be silently ignored on unsupported builds.
Attack Prevented: Supply-chain substitution and reliance on unavailable safeguards.
Prerequisites
An approved deployment owner, the reviewed product version and a disposable isolated test context. Use synthetic data; do not copy production secrets into fixtures.
ClickOps Implementation
- Open NousResearch/hermes-agent → Releases → v0.23.1 and reconcile the source/package identity with your inventory.
- Check the exact executable and the gateway’s deployed revision; preserve a known-good configuration and image digest.
- Treat updates to plugins, skills, MCP servers and container images as separate supply-chain changes. Do not run an upstream installer merely to verify a version.
Code Implementation
The first-party CLI version read previews by default; an explicit opt-in is required to invoke the selected binary.
Code Pack: CLI Script
set +x
set -euo pipefail
[[ $# -eq 0 ]] || { printf '%s\n' 'No positional arguments are accepted.' >&2; exit 2; }
[[ "${HTH_EXECUTE:-0}" == 0 || "${HTH_EXECUTE:-0}" == 1 ]] || exit 2
[[ "${HTH_BINARY:-}" == /* ]] || { printf '%s\n' 'Set HTH_BINARY to an absolute reviewed executable path.' >&2; exit 2; }
PACK_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" && pwd)"
# Default is a preview; opt-in runs only --version, with a timeout and private output.
run_version_read() {
python3 "$PACK_DIR/../../../scripts/lib/agent_hardening.py" version \
--product hermes --version 0.23.1 --binary "$HTH_BINARY" "$@"
}
if [[ "${HTH_EXECUTE:-0}" == 1 ]]; then
run_version_read --execute
else
run_version_read
fi
Validation & Testing
The version pack previews and optionally invokes only --version on an explicit binary. An unrecognized version fails. Confirm the running process separately; a local CLI version does not attest a remote gateway.
Monitoring & Maintenance
Review after product, package, policy or launch changes and at least monthly. Keep the reviewed artifact hash, version, test scope and sanitized result; investigate drift before approving more authority.
Operational Impact
Version upgrades can alter config and approval behavior. Restore only an approved compatible build and config after reviewing state compatibility.
Compliance Mappings
| Framework | Control ID | Relationship |
|---|---|---|
| NIST SP 800-53 Rev. 5 | SI-2, CM-8 | Supports the configuration, least-privilege or evidence objective described above; implementation and assurance require independent review. |
1.2 Pin critical leaves through a protected managed scope
Profile Level: L1 (Crawl)
| Framework | Controls |
|---|---|
| NIST SP 800-53 Rev. 5 | CM-6, AC-3 |
Description
Use administrator-owned /etc/hermes/config.yaml for security-critical leaves, with an administrator-fixed HERMES_MANAGED_DIR launch environment. Managed values override user config and the shell only for the keys they pin. A malformed managed file is logged and ignored, so effective-policy verification is mandatory.
Source: Official configuration/security reference, checked 2026-11-15. Recommendations and test procedures below are HTH’s interpretation of that interface and its limits.
Rationale
Why This Matters:
- Agent-writable configuration can undo a carefully reviewed baseline.
- Higher-precedence sources and fail-open parsing can change effective policy.
Attack Prevented: Configuration tampering and privilege expansion through lower-trust inputs.
Prerequisites
An approved deployment owner, the reviewed product version and a disposable isolated test context. Use synthetic data; do not copy production secrets into fixtures.
ClickOps Implementation
- Open Managed Scope → Where it lives and review the resolved managed directory in your deployment artifact.
- Use your existing deployment mechanism to protect the directory and files from the agent identity. The documented modes are directory 0755 and files 0644; these packs do not change permissions.
- Keep secrets out of the world-readable managed
.env. Pin the bootstrap directory in the supervisor/container, not in a user-editable shell or.env. No graphical fleet-policy workflow was verified for these file settings.
Code Implementation
Candidate output only; review and deploy through your approved change process. The 4.1/4.2 outputs are fragments, while 1.2 is the complete restrictive baseline.
Code Pack: Config
set +x
set -euo pipefail
PACK_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" && pwd)"
python3 "$PACK_DIR/../../../scripts/lib/agent_hardening.py" emit "$@" --product hermes --section all
Validation & Testing
Compare the candidate, then inspect hermes doctor and hermes config privately for the resolved source/pinned keys. Test a harmless conflicting user setting. Verify malformed policy causes an alert and admission refusal in your launcher; Hermes itself continues without that policy.
Monitoring & Maintenance
Review after product, package, policy or launch changes and at least monthly. Keep the reviewed artifact hash, version, test scope and sanitized result; investigate drift before approving more authority.
Operational Impact
Managed leaves constrain existing workflows but do not freeze unspecified keys or arbitrary subprocess behavior. Restore an approved managed artifact and restart to recover.
Compliance Mappings
| Framework | Control ID | Relationship |
|---|---|---|
| NIST SP 800-53 Rev. 5 | CM-6, AC-3 | Supports the configuration, least-privilege or evidence objective described above; implementation and assurance require independent review. |
2. Execution and Network Boundaries
2.1 Isolate the whole agent rather than only terminal tools
Profile Level: L2 (Walk)
| Framework | Controls |
|---|---|
| NIST SP 800-53 Rev. 5 | AC-6, SC-39 |
Description
Hermes is a single-tenant personal agent. Its SECURITY.md treats OS isolation as the containment boundary. A Docker terminal backend confines shell/file tools, while the Python agent, execute_code host subprocesses, MCP, plugins, hooks and skill loading can remain outside that backend.
Source: Official configuration/security reference, checked 2026-11-15. Recommendations and test procedures below are HTH’s interpretation of that interface and its limits.
Rationale
Why This Matters:
- Prompt injection can steer any capability the process can reach.
- Tool-level safeguards do not contain every extension or subprocess path.
Attack Prevented: Host credential theft and cross-workspace effects from untrusted content.
Prerequisites
An approved deployment owner, the reviewed product version and a disposable isolated test context. Use synthetic data; do not copy production secrets into fixtures.
ClickOps Implementation
- Read Security Policy → Whole-process wrapping before accepting web/email/shared-channel input.
- Place the complete Hermes process tree, including the selected gateway/backend, in an approved container/VM/sandbox; expose only task mounts and approved network services. Do not mount a developer home, Docker socket or production credential store.
- Review the launch preview. Isolating only a client does not isolate a shared gateway executing elsewhere. The candidate’s
terminal.backend: dockeris only an additional shell/file boundary; it does not create whole-agent isolation, and backend images need their own approved version/digest.
Code Implementation
This produces a plan only. It does not create or attest the external isolation boundary.
Code Pack: Config
set +x
set -euo pipefail
PACK_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" && pwd)"
python3 "$PACK_DIR/../../../scripts/lib/agent_hardening.py" launch "$@" --product hermes
Validation & Testing
Use disposable harmless files to test host and network reachability through terminal, code execution and a permitted MCP/plugin path independently. Verify the external boundary rather than inferring it from terminal backend selection.
Monitoring & Maintenance
Review after product, package, policy or launch changes and at least monthly. Keep the reviewed artifact hash, version, test scope and sanitized result; investigate drift before approving more authority.
Operational Impact
Whole-process isolation may require redesigning credentials and state mounts. Recreate the approved environment instead of broadening host access to make a failing tool work.
Compliance Mappings
| Framework | Control ID | Relationship |
|---|---|---|
| NIST SP 800-53 Rev. 5 | AC-6, SC-39 | Supports the configuration, least-privilege or evidence objective described above; implementation and assurance require independent review. |
2.2 Deny open gateway and unauthenticated API access
Profile Level: L1 (Crawl)
| Framework | Controls |
|---|---|
| NIST SP 800-53 Rev. 5 | AC-3, SC-7 |
Description
Keep gateway access restricted to exact platform users or approved DM pairing. Never enable global allow-all on a work deployment. The API server exposes agent tools including terminal execution and requires API_SERVER_KEY even on loopback; narrow browser origins and use strong public-dashboard authentication.
Source: Official configuration/security reference, checked 2026-11-15. Recommendations and test procedures below are HTH’s interpretation of that interface and its limits.
Rationale
Why This Matters:
- Agent listeners and outbound services expose real execution authority.
- A proxy or loopback default alone does not prove all traffic is constrained.
Attack Prevented: Unauthorized remote actions and exfiltration through overlooked network paths.
Prerequisites
An approved deployment owner, the reviewed product version and a disposable isolated test context. Use synthetic data; do not copy production secrets into fixtures.
ClickOps Implementation
- Review Security → User authorization (gateway): retain platform
*_ALLOWED_USERS/pairing andGATEWAY_ALLOW_ALL_USERS=false; also check bridgedgateway.allow_all_usersvalues and secondary-profile environments. - If the API is required, keep
API_SERVER_HOST=127.0.0.1, inject a strongAPI_SERVER_KEYand explicitly restrictAPI_SERVER_CORS_ORIGINS. Do not enable the API just to test this guide. - For a public dashboard follow Web Dashboard → Authentication with OAuth/OIDC. Username/password is for trusted networks; the agent API key and dashboard login are different boundaries.
Code Implementation
Verification pack: a bounded local check, not an enforcement operation or effective-policy attestation. See the pack README for exact inputs, output and omissions.
Code Pack: Config
set +x
set -euo pipefail
PACK_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" && pwd)"
python3 "$PACK_DIR/../../../scripts/lib/agent_hardening.py" environment "$@" --product hermes --scope network
Validation & Testing
The environment audit flags open gateway, missing enabled-API key, non-loopback bind and wildcard CORS in its own environment only. In an authorized test context, deny an unknown user and unauthenticated API request, and inspect the real listeners/profile environments.
Monitoring & Maintenance
Review after product, package, policy or launch changes and at least monthly. Keep the reviewed artifact hash, version, test scope and sanitized result; investigate drift before approving more authority.
Operational Impact
Narrowing access can disconnect bots and clients. Keep a separate operator access path and redeploy an approved roster; never recover by turning on allow-all.
Compliance Mappings
| Framework | Control ID | Relationship |
|---|---|---|
| NIST SP 800-53 Rev. 5 | AC-3, SC-7 | Supports the configuration, least-privilege or evidence objective described above; implementation and assurance require independent review. |
3. Credentials and Integrations
3.1 Minimize forwarded credentials and skill-declared access
Profile Level: L1 (Crawl)
| Framework | Controls |
|---|---|
| NIST SP 800-53 Rev. 5 | AC-6, IA-5 |
Description
Keep terminal.env_passthrough, terminal.docker_forward_env and terminal.credential_files empty unless a task explicitly requires them. Skill declarations can add required environment variables and credential files to sandbox access, so empty explicit arrays alone do not prove credentials are withheld.
Source: Official configuration/security reference, checked 2026-11-15. Recommendations and test procedures below are HTH’s interpretation of that interface and its limits.
Rationale
Why This Matters:
- Ambient credentials enlarge the impact of one hostile tool or resource.
- Model selection and a successful login do not prove approved account scope.
Attack Prevented: Credential misuse and unauthorized provider or internal-service access.
Prerequisites
An approved deployment owner, the reviewed product version and a disposable isolated test context. Use synthetic data; do not copy production secrets into fixtures.
ClickOps Implementation
- Inspect Security → Skill environment passthrough / Credential files and review every enabled skill’s
required_environment_variablesandrequired_credential_files. - Keep the candidate’s forwarding arrays empty; supply task-scoped, short-lived credentials only through an approved injection path. Do not place sensitive keys in managed
.env. - Review MCP environment overrides, host subprocess reach, SSH-agent/socket access and external secret integrations. The calling-process audit is only a bounded signal, not a complete credential inventory.
Code Implementation
Verification pack: a bounded local check, not an enforcement operation or effective-policy attestation. See the pack README for exact inputs, output and omissions.
Code Pack: Config
set +x
set -euo pipefail
PACK_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" && pwd)"
python3 "$PACK_DIR/../../../scripts/lib/agent_hardening.py" environment "$@" --product hermes --scope credentials
Validation & Testing
Use dummy markers to verify a denied credential is unavailable through shell, execute_code and MCP independently. Never print real environments, auth files or secret values as evidence.
Monitoring & Maintenance
Review after product, package, policy or launch changes and at least monthly. Keep the reviewed artifact hash, version, test scope and sanitized result; investigate drift before approving more authority.
Operational Impact
Credential minimization can disable integrations. Grant the smallest task-specific access and repeat the isolated negative test after skill or plugin changes.
Compliance Mappings
| Framework | Control ID | Relationship |
|---|---|---|
| NIST SP 800-53 Rev. 5 | AC-6, IA-5 | Supports the configuration, least-privilege or evidence objective described above; implementation and assurance require independent review. |
3.2 Review executable plugins, MCP and persistent automation
Profile Level: L2 (Walk)
| Framework | Controls |
|---|---|
| NIST SP 800-53 Rev. 5 | CM-7, SR-3 |
Description
Treat user plugins, hook handlers, skills, MCP commands and scheduled jobs as persistent executable capabilities. Terminal-backend isolation does not contain their host code paths. Inventory and approve their source, version, credentials and triggers before enabling them.
Source: Official configuration/security reference, checked 2026-11-15. Recommendations and test procedures below are HTH’s interpretation of that interface and its limits.
Rationale
Why This Matters:
- Resources can introduce executable code, instructions and persistent access.
- Workspace inspection alone misses user, profile and system discovery roots.
Attack Prevented: Supply-chain execution and unauthorized third-party tool access.
Prerequisites
An approved deployment owner, the reviewed product version and a disposable isolated test context. Use synthetic data; do not copy production secrets into fixtures.
ClickOps Implementation
- Review the user Hermes home, plugin directory, profile skills, MCP definitions and scheduled jobs in your approved editor. Current gateway architecture discovers user plugins once, not per-project plugins.
- Keep unapproved plugins disabled and inspect startup hooks and skill dependencies. A workspace-only inventory does not enumerate the actual user home or gateway state.
- For MCP use the official reference’s trust/credential controls; retain approval for untrusted write-capable tools. Do not represent a tool label or trust setting as process containment.
Code Implementation
Verification pack: a bounded local check, not an enforcement operation or effective-policy attestation. See the pack README for exact inputs, output and omissions.
Code Pack: Config
set +x
set -euo pipefail
PACK_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" && pwd)"
python3 "$PACK_DIR/../../../scripts/lib/agent_hardening.py" inventory "$@" --product hermes
Validation & Testing
Compare the approved manifest with the effective tool/plugin/job inventory after restart. Test only harmless local fixtures; never connect to an unapproved remote MCP server to collect evidence.
Monitoring & Maintenance
Review after product, package, policy or launch changes and at least monthly. Keep the reviewed artifact hash, version, test scope and sanitized result; investigate drift before approving more authority.
Operational Impact
Disabling a persistent integration can interrupt jobs. Suspend through the supported operator workflow and preserve a reviewed manifest; removing a workspace file does not revoke its upstream grants.
Compliance Mappings
| Framework | Control ID | Relationship |
|---|---|---|
| NIST SP 800-53 Rev. 5 | CM-7, SR-3 | Supports the configuration, least-privilege or evidence objective described above; implementation and assurance require independent review. |
4. Tool and Data Restrictions
4.1 Keep flagged-command approvals manual and headless paths denied
Profile Level: L1 (Crawl)
| Framework | Controls |
|---|---|
| NIST SP 800-53 Rev. 5 | AC-3, CM-7 |
Description
Set approvals.mode: manual, headless cron/single-query/unattended policies to deny, and narrow approvals.approvers to the requester/platform administrators with an empty list. Manual mode governs detected dangerous commands, not every tool call. Container backends can skip dangerous-command checks; the OS boundary remains essential.
Source: Official configuration/security reference, checked 2026-11-15. Recommendations and test procedures below are HTH’s interpretation of that interface and its limits.
Rationale
Why This Matters:
- Alternate execution paths can bypass a narrowly configured tool rule.
- Headless behavior and explicit overrides differ from interactive defaults.
Attack Prevented: Unapproved mutation, execution and delegation by an injected instruction.
Prerequisites
An approved deployment owner, the reviewed product version and a disposable isolated test context. Use synthetic data; do not copy production secrets into fixtures.
ClickOps Implementation
- In the protected candidate, review
approvalsand preservemcp_reload_confirmanddestructive_slash_confirmas true. - Keep
command_allowlistempty. Allowlist rules can override unattended denial; removing a remembered rule requires restart to clear in-memory decisions. - Reject unapproved
--yolo,/yolo,HERMES_YOLO_MODEandapprovals.mode: off. Review resumed sessions because session YOLO state can persist.
Code Implementation
Candidate output only; review and deploy through your approved change process. The 4.1/4.2 outputs are fragments, while 1.2 is the complete restrictive baseline.
Code Pack: Config
set +x
set -euo pipefail
PACK_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" && pwd)"
python3 "$PACK_DIR/../../../scripts/lib/agent_hardening.py" emit "$@" --product hermes --section 4.1
Validation & Testing
With harmless fixtures, test a flagged local command, a headless deny and approval by an unapproved chat participant. Do not extrapolate those results to all commands or to a Docker terminal. Restart after approval-memory changes and repeat.
Monitoring & Maintenance
Review after product, package, policy or launch changes and at least monthly. Keep the reviewed artifact hash, version, test scope and sanitized result; investigate drift before approving more authority.
Operational Impact
Headless tasks that need dangerous operations will fail and need redesign or a scoped operator action. Do not restore availability by enabling blanket approval.
Compliance Mappings
| Framework | Control ID | Relationship |
|---|---|---|
| NIST SP 800-53 Rev. 5 | AC-3, CM-7 | Supports the configuration, least-privilege or evidence objective described above; implementation and assurance require independent review. |
4.2 Redact output and protect session and checkpoint data
Profile Level: L1 (Crawl)
| Framework | Controls |
|---|---|
| NIST SP 800-53 Rev. 5 | AC-4, AU-9 |
Description
Retain security.redact_secrets: true and enable checkpoints for recoverable work. Redaction is a heuristic for output, not a DLP or credential boundary. Checkpoints, conversations, memory and skills are local data that can themselves contain sensitive source and tool output.
Source: Official configuration/security reference, checked 2026-11-15. Recommendations and test procedures below are HTH’s interpretation of that interface and its limits.
Rationale
Why This Matters:
- Source, tokens and session evidence can escape through several paths.
- Convenience features and heuristic filtering leave residual exposure.
Attack Prevented: Sensitive-data disclosure and reuse of retained authority or context.
Prerequisites
An approved deployment owner, the reviewed product version and a disposable isolated test context. Use synthetic data; do not copy production secrets into fixtures.
ClickOps Implementation
- Pin output redaction in the managed candidate and review the private Hermes state/checkpoint storage available to the process.
- Use synthetic token-shaped fixtures to evaluate the redaction paths you actually use. Keep production keys outside prompts, files and shared evidence.
- Enable checkpoints only in the approved state location; preview
/rollback diffbefore any deliberate recovery.HERMES_WRITE_SAFE_ROOTconstrains write_file/patch only and may block Hermes state writes unless the required state root is included.
Code Implementation
Candidate output only; review and deploy through your approved change process. The 4.1/4.2 outputs are fragments, while 1.2 is the complete restrictive baseline.
Code Pack: Config
set +x
set -euo pipefail
PACK_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" && pwd)"
python3 "$PACK_DIR/../../../scripts/lib/agent_hardening.py" emit "$@" --product hermes --section 4.2
Validation & Testing
The candidate audit checks redaction/checkpoint values, not stored-data content, retention or successful restoration. In an authorized disposable environment, test a synthetic redaction and recovery; preserve the untouched original fixture for comparison.
Monitoring & Maintenance
Review after product, package, policy or launch changes and at least monthly. Keep the reviewed artifact hash, version, test scope and sanitized result; investigate drift before approving more authority.
Operational Impact
Checkpoints consume storage and may retain deleted secrets. Apply a private retention policy and use reviewed recovery rather than automatic replay of old config.
Compliance Mappings
| Framework | Control ID | Relationship |
|---|---|---|
| NIST SP 800-53 Rev. 5 | AC-4, AU-9 | Supports the configuration, least-privilege or evidence objective described above; implementation and assurance require independent review. |
5. Monitoring and Verification
5.1 Check candidate drift and prove managed enforcement
Profile Level: L1 (Crawl)
| Framework | Controls |
|---|---|
| NIST SP 800-53 Rev. 5 | CM-6, CA-7 |
Description
Treat missing or ignored managed policy as an admission failure in your deployment wrapper. The offline audit checks an explicit candidate file against the reviewed baseline; it does not emulate managed/user/environment merging or inspect a running gateway.
Source: Official configuration/security reference, checked 2026-11-15. Recommendations and test procedures below are HTH’s interpretation of that interface and its limits.
Rationale
Why This Matters:
- Syntactically valid configuration can still be ignored or overridden.
- A file-only check cannot attest a running process or its external boundary.
Attack Prevented: Silent policy drift and false confidence from incomplete verification.
Prerequisites
An approved deployment owner, the reviewed product version and a disposable isolated test context. Use synthetic data; do not copy production secrets into fixtures.
ClickOps Implementation
- Run the candidate audit before deployment and investigate every extra, missing or incorrectly typed key.
- After restart, privately review
hermes doctorand the managed header/pinned keys inhermes config. - Reconcile actual bootstrap paths, enabled resources and environment bridges with the approved deployment. Alert on parse warnings and unexpected changes rather than accepting startup as proof.
Code Implementation
Verification pack: a bounded local check, not an enforcement operation or effective-policy attestation. See the pack README for exact inputs, output and omissions.
Code Pack: Config
set +x
set -euo pipefail
PACK_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" && pwd)"
python3 "$PACK_DIR/../../../scripts/lib/agent_hardening.py" audit "$@" --product hermes
Validation & Testing
Offline fixtures include malformed YAML, duplicate keys, aliases, unsafe types, nonempty command allowlists and widened approvers. Then test effective conflicting user settings in the isolated deployment; no such live proof was performed for this guide.
Monitoring & Maintenance
Review after product, package, policy or launch changes and at least monthly. Keep the reviewed artifact hash, version, test scope and sanitized result; investigate drift before approving more authority.
Operational Impact
Strict comparison rejects customizations until reviewed. Preserve the approved baseline and update source evidence/tests with each accepted exception.
Compliance Mappings
| Framework | Control ID | Relationship |
|---|---|---|
| NIST SP 800-53 Rev. 5 | CM-6, CA-7 | Supports the configuration, least-privilege or evidence objective described above; implementation and assurance require independent review. |
5.2 Preserve private diagnostics and rehearse containment
Profile Level: L1 (Crawl)
| Framework | Controls |
|---|---|
| NIST SP 800-53 Rev. 5 | AU-6, AU-9 |
Description
Record version, managed sources, gateway authorization decisions and resource changes without publishing tokens or transcripts. The fixed native doctor command is an operator diagnostic; its zero exit code is not a security verdict.
Source: Official configuration/security reference, checked 2026-11-15. Recommendations and test procedures below are HTH’s interpretation of that interface and its limits.
Rationale
Why This Matters:
- Diagnostics and session evidence may themselves contain secrets.
- Recovery must account for retained processes, grants and external effects.
Attack Prevented: Persistent unauthorized access and secondary disclosure during response.
Prerequisites
An approved deployment owner, the reviewed product version and a disposable isolated test context. Use synthetic data; do not copy production secrets into fixtures.
ClickOps Implementation
- Preview the native diagnostic pack and execute only in the approved test context; its output is intentionally suppressed.
- Review raw doctor/config diagnostics privately and retain only sanitized rule outcomes, source/release identifiers and test dates.
- For suspected compromise, use your operator response procedure to stop affected agent/gateway processes and persistent jobs, inspect whole-process access and revoke upstream grants as appropriate. No response action is automated here.
Code Implementation
Native reads preview by default. --execute is available only for the fixed diagnostic commands, never remediation; completion is not reported as a policy pass.
Code Pack: CLI Script
set +x
set -euo pipefail
PACK_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" && pwd)"
# Default prints fixed argv; --execute opts into native reads only, never remediation.
python3 "$PACK_DIR/../../../scripts/lib/agent_hardening.py" diagnostics "$@" --product hermes
Validation & Testing
Rehearse with harmless fixtures and a synthetic unapproved user. Verify access is refused after containment and that retained jobs/sessions do not resume outside the approved boundary.
Monitoring & Maintenance
Review after product, package, policy or launch changes and at least monthly. Keep the reviewed artifact hash, version, test scope and sanitized result; investigate drift before approving more authority.
Operational Impact
Stopping the gateway interrupts every dependent channel/profile. Coordinate recovery from the approved manifest and verify authorization again before reopening access.
Compliance Mappings
| Framework | Control ID | Relationship |
|---|---|---|
| NIST SP 800-53 Rev. 5 | AU-6, AU-9 | Supports the configuration, least-privilege or evidence objective described above; implementation and assurance require independent review. |
6. Compliance Quick Reference
These are suggested control relationships against the NIST SP 800-53 Rev. 5 catalog, not a product benchmark or certification. Bounded searches did not identify a product-specific CIS/SCuBA baseline; DISA coverage remains unverified. No invented benchmark IDs or edition guarantees are used.
| Objective | Guide controls | Evidence to retain privately |
|---|---|---|
| Asset identity and configuration | 1.1, 1.2, 5.1 | Exact build, approved artifact hash, actual source/precedence and drift result |
| Least privilege and information flow | 2.1–4.2 | Task mounts/credentials, external network/OS denies, effective resources and harmless negative tests |
| Monitoring and response | 5.2 | Sanitized diagnostic outcomes, grant/session review and recovery test |
7. Appendices
Deployment and capability matrix
| Surface | Applicability and limits |
|---|---|
| Reviewed product release | 0.23.1; other versions require explicit source reconciliation |
| Local configuration and CLI | Candidate generation/comparison and fixed opt-in native reads; no automatic deployment |
| Whole-process boundary | Supplied by your approved external environment; the launch plan is not a sandbox |
| Tenant identity/SSO | Apply upstream provider/account controls separately; no unverified tenant API is fabricated |
| API/SDK/Terraform/SIEM | See the surface census; available operational APIs are not assumed to expose policy administration |
| Validation maturity | ai-drafted; source, schema and fixture checks are not live or human review |
Integration Risk Assessment Matrix
| Integration | Principal exposure | Admission requirement |
|---|---|---|
| Model provider | Source and inference credentials | Approved endpoint/account, retention terms and restricted credential/egress scope |
| MCP server | Tool authority, remote data and subprocess credentials | Reviewed command/endpoint, exact tool scope and independent process boundary |
| Extension/plugin/skill | Startup code or instructions with process authority | Reviewed source/version, discovery roots, credentials and persistence |
| Browser/remote API | Authenticated actions or remote execution | Task-specific account, restricted ingress and explicit capability admission |
References
- First-party reference
- First-party reference
- First-party reference
- First-party reference
- First-party reference
- First-party reference
- First-party reference
- Source coverage, automation census and release pins
- Code Pack usage and evidence limits
Changelog
| Date | Version | Maturity | Changes | Author |
|---|---|---|---|---|
| 2026-11-15 | 0.1.0 | ai-drafted | Initial source-backed guide and version-gated candidate, audit, inventory and native-read packs; offline fixtures only. | Codex (GPT-6) |
Contributing
Report corrections with the exact release, first-party source and a sanitized reproduction. Follow CONTRIBUTING.md. Keep credentials and private agent/tenant evidence out of public issues and pull requests.