🛡️ TrustMyAgent - Security Agent

Stateless security posture monitoring for AI agents.

TrustMyAgent runs 41 read-only security checks across 14 domains on the host machine, calculates a trust score (0-100), and optionally reports results to a public Trust Center dashboard. It is fully stateless — no files are written, no state is persisted, and no background processes are installed without explicit user consent.

Open source: All check logic is visible in run.py and checks/*.json. The full source code is at github.com/Anecdotes-Yair/trust-my-agent-ai.

Setup

When the user asks to set up TrustMyAgent, walk them through the following steps. Present each step to the user and wait for their approval before executing. Do not run commands silently or skip user confirmation.

Step 1: Verify Python 3

Check that Python 3 is available:

python3 --version

If Python 3 is not installed, ask the user before installing it:

• macOS: brew install python3
• Linux: sudo apt-get install python3 or sudo apk add python3

Step 2: Run a dry-run assessment

Run the security assessment in dry-run mode first so the user can review the results before any data leaves the machine:

python3 {baseDir}/run.py --dry-run

This runs all 41 checks and prints the full telemetry payload that would be sent — but sends nothing. Show the user the results and let them review:

• The trust score and tier
• Which checks passed and failed
• The exact JSON payload that would be transmitted

Ask the user: "Would you like to send this report to the Trust Center dashboard, run in local-only mode (no network), or adjust any settings?"

Step 3: Run the full assessment (with user consent)

If the user approves sending telemetry:

python3 {baseDir}/run.py

If the user prefers local-only mode (no telemetry sent):

python3 {baseDir}/run.py --local-only

Step 4: Schedule recurring assessments (optional)

Ask the user if they want to schedule automatic assessments. If they agree, suggest a cron job. The command varies by agent platform:

OpenClaw agents:

openclaw cron add --name "🛡️ TrustMyAgent" --cron "*/15 * * * *" --session agent:security:main --message "Run security assessment: python3 {baseDir}/run.py"

Other agents / standalone:

# Add to system crontab (runs every 15 minutes)
(crontab -l 2>/dev/null; echo "*/15 * * * * python3 {baseDir}/run.py --quiet") | crontab -

The user can choose any interval or skip scheduling entirely.

What Data Is Sent

When telemetry is enabled (the default), the following data is sent via HTTPS POST to https://www.trustmyagent.ai/api/telemetry:

What is NOT sent:

• No file contents, paths, or directory listings
• No environment variable values (only whether secret-like patterns exist)
• No process names, PIDs, or command lines
• No network traffic, IP addresses, or hostnames
• No credentials, tokens, or API keys
• No conversation transcripts or user data

The telemetry endpoint and all check logic are open source. You can verify exactly what is transmitted by using --dry-run mode.

Opting out of telemetry

Use --local-only to run all checks without any network calls:

python3 {baseDir}/run.py --local-only

This gives you the full security assessment locally without sending anything.

How It Works

1. run.py executes on the host — triggered manually, by cron, or by agent heartbeat
2. 41 security checks run using bash commands and Python sensors (all read-only)
3. Trust score is calculated (0-100) based on pass/fail results and severity weighting
4. Results are displayed locally in the terminal
5. (Optional) Telemetry is sent to the Trust Center dashboard via HTTPS

No files are written locally. No state is persisted on the agent machine.

Security Domains

Check Types

Bash checks (20 checks)

Defined in checks/openclaw_checks.json. Each check runs a shell command and evaluates the output against a pass_condition (equals, contains, not_contains, exit_code_zero, etc.).

Python/Message-based checks (21 checks)

Defined in checks/message_checks.json and checks/nodes_media_checks.json. These are programmatic sensors that analyze secrets, session transcripts, MCP configs, skill manifests, and more.

Platform Support

Checks auto-detect macOS vs Linux and use platform-appropriate commands. Checks can declare "platforms": ["linux"] to be gracefully skipped on unsupported platforms.

Trust Tiers

Any critical-severity failure caps the score at 49 (UNTRUSTED). Three or more high-severity failures cap at 69 (LOW).

Command Line Options

Configuration

Files

Agent/
├── SKILL.md                        # This file
├── run.py                          # Main entry point (stateless runner)
└── checks/
    ├── openclaw_checks.json        # 20 bash-based security checks
    ├── message_checks.json         # 10 Python-based message/secret sensors
    ├── nodes_media_checks.json     # 11 infrastructure checks
    └── detection_kb.json           # Risk descriptions and remediation guidance

Architecture

┌─────────────────┐                                 ┌──────────────────┐
│   Agent Host     │      POST /api/telemetry        │ 🛡️ TrustMyAgent  │
│                  │  ────────────────────────────►   │  Server           │
│  run.py          │  (only when telemetry enabled)  │  (Cloudflare)    │
│  ├─ bash checks  │                                 │  ├─ R2 storage   │
│  └─ python checks│                                 │  ├─ agents index │
│                  │                                 │  └─ trend history│
│  (no local state)│                                 │                  │
└─────────────────┘                                  └──────────────────┘
                                                            │
                                                     trust-center.html
                                                     (public dashboard)

Privacy & Trust

• Open source: All code is MIT-licensed and publicly auditable at github.com/Anecdotes-Yair/trust-my-agent-ai
• Stateless: No files written, no state persisted, no background processes installed without consent
• Opt-in telemetry: Use --local-only to run entirely offline, or --dry-run to preview before sending
• No secrets transmitted: Checks detect the presence of issues, never transmit actual secret values
• Transparent payload: The --dry-run flag shows the exact JSON that would be sent
• Server: Operated by Anecdotes AI, a GRC (Governance, Risk, Compliance) company. Server code is at github.com/Anecdotes-Yair/trust-my-agent-ai-website

Credits

Built by Anecdotes AI for the AI agent ecosystem.