TrustMyAgent

🛡️ TrustMyAgent - Security posture monitoring for AI agents. Runs 41 stateless checks across 14 domains and calculates a trust score (0-100). Supports local-only mode (no network calls) and dry-run mode (preview before sending).

🛡️ 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:

bash

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:

bash

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:

bash

python3 {baseDir}/run.py

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

bash

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:

bash

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

Other agents / standalone:

bash

# 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:

Field	Example	Purpose
agent.id	sha256(hostname)	Unique identifier (derived from hostname hash, not the hostname itself)
agent.name	"My Agent"	Display name (from IDENTITY.md or env var)
agent.platform	"darwin"	OS type (darwin/linux)
agent.detected_env	"macos_arm64"	Runtime environment label
posture.trust_tier	"HIGH"	Computed trust tier
posture.overall_score	92	Numeric score 0-100
results[]	{check_id, passed, status}	Per-check pass/fail status
detections[]	{check_id, severity, risk}	Failed checks with risk context

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:

bash

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

Domain	Checks	Focus
Physical Environment	PHY-001 to PHY-005	Disk encryption, container isolation, non-root execution
Network	NET-001 to NET-005	Dangerous ports, TLS/SSL, DNS, certificates
Secrets	SEC-001 to SEC-005, MSG-005	Env var secrets, cloud creds, private keys, conversation leaks
Code	COD-001 to COD-004	Git security, no secrets in repos
Logs	LOG-001 to LOG-004	System logging, audit readiness
Skills	SKL-001 to SKL-005, MSG-001, MSG-003	Skill manifests, MCP server trust
Integrity	INT-001 to INT-005, MSG-002, MSG-006	Backdoors, browser abuse, suspicious tool calls, URL reputation
Social Guards	SOC-001 to SOC-006, MSG-004	Action logging, session transparency, Moltbook integrity, owner reputation
Incident Prevention	INC-001 to INC-005	Process spawning, system load, port scanning
Node Security	NODE-001 to NODE-005	Remote execution approval, token permissions, exec allowlists
Media Security	MEDIA-002 to MEDIA-003	Temp directory permissions, file type validation
Gateway Security	GATEWAY-001 to GATEWAY-002	Binding address, authentication
Identity Security	IDENTITY-001 to IDENTITY-002	DM pairing allowlist, group chat allowlist
SubAgent Security	SUBAGENT-001 to SUBAGENT-002	Concurrency limits, target allowlists

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

Tier	Score	Label
HIGH	90-100	Ready for Business
MEDIUM	70-89	Needs Review
LOW	50-69	Elevated Risk
UNTRUSTED	0-49	Critical Security Gaps

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

Command Line Options

Flag	Description
--checks, -c	Path to custom checks JSON file
--timeout, -t	Timeout per check in seconds (default: 30)
--quiet, -q	Minimal output
--json, -j	Output structured JSON to stdout
--dry-run	Run all checks and display the telemetry payload, but do not send it
--local-only	Run all checks locally without any network calls
--no-notify	Skip agent notifications for detections

Configuration

Source	Description	Default
IDENTITY.md	Agent display name (read from # Name section)	"Agent"
OPENCLAW_AGENT_NAME env var	Overrides IDENTITY.md name	—
OPENCLAW_AGENT_ID env var	Agent identifier	SHA256 of hostname
TRUSTMYAGENT_TELEMETRY_URL env var	Server endpoint	https://www.trustmyagent.ai/api/telemetry

Files

text

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

text

┌─────────────────┐                                 ┌──────────────────┐
│   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.