Back to skill

Security audit

VexPath Skill Pack

Security checks for vulnerabilities and agentic risk

Overview

VexPath mostly matches its business-operations purpose, but it asks for broad email access, stores mailbox passwords insecurely, installs persistent monitoring instructions, and includes risky provisioning scripts.

Review this carefully before installing. Do not paste mailbox passwords into command lines or chat, do not run the VPS curl-to-bash provisioner, and only enable heartbeat monitoring after confirming exactly which inboxes, calendars, CRM records, and memory files it may access and retain. Prefer OAuth or a keychain/password-manager setup, pin installation dependencies, and avoid deploying the workspace SOUL.md/HEARTBEAT.md files unless you explicitly want persistent VEX behavior.

Vulnerability Patterns
  • Skill Instruction HijackingAlters the agent's session goals or safety constraints when the skill loads
  • Remote Payload Retrieval and ExecutionFetches external code whose behavior can change after review
  • Insecure DependenciesIntroduces malicious components through unsafe dependency sources
  • Insecure Skill Coding PracticesFinds exploitable flaws such as hardcoded secrets or command injection
  • Agent Memory PoisoningWrites attacker-controlled rules into memory that affect later sessions
Findings (6)

T03 · Remote Payload Retrieval and Execution

Error
Location
scripts/provision-vps.sh:47
Finding

Remote NodeSource Bootstrap Script Executed Directly as Root

Content
View full analysis
/dev/null || [[ $(node -v | sed 's/v//' | cut -d. -f1) -lt 20 ]]; then log "Installing Node.js 22..." curl -fsSL https://deb.nodesource.com/setup_22.x | bash - apt-get install -y nodejs fi ``` ### Technical Analysis The provisioner downloads a mutable shell script from `deb.nodesource.com` and immediately pipes it into `bash`. There is no version pinning, checksum validation, signature verification, or opportunity to inspect the downloaded content before execution. The following `apt-get` command demonstrates that the provisioner is expected to run with root privileges. Consequently, the remotely supplied bootstrap script also executes as root. HTTPS protects the connection in transit but does not protect against a compromised upstream server, compromised publishing account, malicious upstream change, DNS/TLS trust compromise, or unexpected behavior introduced after this project was audited. Executing a remote repository configuration script is not the minimum privilege necessary to install Node.js. A signed package repository can be configured using explicit, reviewable steps without running mutable network content as shell code. ### Attack Path 1. An attacker compromises the NodeSource distribution endpoint, its publishing process, or another component in its delivery chain. 2. The attacker modifies `setup_22.x` to include arbitrary commands. 3. An administrator runs `provision-vps.sh` with root privileges. 4. `curl` retrieves the modified script and passes it directly to `bash`. 5. The payload executes with root privileges before Node.js installation continues. 6. The payload can alter the operating system, steal secrets, create privileged accounts, or establish persistence. ### Impact Assessment Successf ...[truncated 446 chars]
Remediation
View remediation

T03 · Remote Payload Retrieval and Execution

Error
Location
scripts/provision-vps.sh:5
Finding

Installation Instructions Execute an Unspecified Hosted Script with Email Credentials

Content
View full analysis
/provision-vps.sh | bash -s -- \ # "" "" [business-type] ``` The downloaded script later consumes the supplied secret: ```bash CLIENT_EMAIL="${1:-}" DISPLAY_NAME="${2:-}" APP_PASSWORD="${3:-}" BIZ_TYPE="${4:-service}" ``` ### Technical Analysis The recommended installation method executes shell content returned by an unspecified hosted URL. The effective payload is therefore not fixed to the code that was audited. A malicious or replaced hosted script can immediately access the email address and app password supplied as arguments. The script is intended to provision system packages and globally install software, so users are likely to execute it as root. This combines two sensitive capabilities in one operation: - Arbitrary execution of mutable remote content. - Direct availability of a valid mailbox credential to that content. Using a placeholder such as `` provides no verifiable trust anchor, release identity, signature, or immutable revision. ### Attack Path 1. A user substitutes a compromised, mistyped, or attacker-controlled URL into the documented command, or the legitimate hosting account is compromised. 2. The remote endpoint returns a modified provisioning script. 3. The returned content is executed immediately by `bash`, commonly as root. 4. The malicious script reads `$1`, `$2`, and `$3`, including the mailbox app password. 5. It transmits the credential to an attacker or uses it to access the mailbox. 6. It can also use its root execution context to compromise the VPS and persist after installation. ### Impact Assessment The attack can provide complete control of the target VPS and unauthorized IMAP/SMTP access to the client mailbox. Depend ...[truncated 253 chars]
Remediation
View remediation

T01 · Skill Instruction Hijacking

Error
Location
scripts/deploy-client.sh:70
Finding

Deployment Replaces Persistent Agent Identity and Installs Recurring Inbox Instructions

Content
View full analysis
Remediation
View remediation

T09 · Insecure Skill Coding Practices

Error
Location
scripts/setup-email.sh:72
Finding

Email Passwords Are Passed in Process Arguments and Stored in Plaintext

Content
View full analysis
"" "" # Example: ./setup-email.sh you@gmail.com "Jane Smith" "abcd efgh ijkl mnop" ``` ```bash EMAIL="$1" DISPLAY_NAME="$2" APP_PASSWORD="$(echo "$3" | tr -d ' ')" DOMAIN="${EMAIL##*@}" ``` ```bash backend.login = "${EMAIL}" backend.encryption.type = "tls" backend.auth.type = "password" backend.auth.raw = "${APP_PASSWORD}" message.send.backend.type = "smtp" message.send.backend.host = "${SMTP_HOST}" message.send.backend.port = ${SMTP_PORT} message.send.backend.login = "${EMAIL}" message.send.backend.encryption.type = "tls" message.send.backend.auth.type = "password" message.send.backend.auth.raw = "${APP_PASSWORD}" ``` The generated file is restricted afterward: ```bash chmod 600 "$CONFIG_FILE" ``` ### Technical Analysis The setup interface requires users to provide the app password as a command-line argument. This may expose the secret through shell history, process inspection, terminal recording, audit tooling, support logs, or command telemetry. The script then interpolates the password directly into `config.toml`. Permission mode `600` reduces access to the owning account, but it does not encrypt the credential or isolate it from other processes running as that user. The configuration can also be copied into backups because existing configurations are duplicated before replacement. The documentation additionally recommends direct account passwords for Hostinger accounts and characterizes plaintext local storage as acceptable. Direct account passwords generally have a larger impact than revocable, application-specific credentials. ### Atta ...[truncated 1003 chars]
Remediation
View remediation

T08 · Insecure Dependencies

Error
Location
scripts/provision-vps.sh:61
Finding

Unpinned Global npm Packages Are Installed During Privileged Provisioning

Content
View full analysis
/dev/null; then log "Installing OpenClaw..." npm install -g openclaw else log "OpenClaw already installed: $(openclaw --version 2>/dev/null || echo 'unknown')" fi ``` ```bash else log "ClawHub not found — installing npm + clawhub..." npm install -g clawhub 2>/dev/null || true if command -v clawhub &>/dev/null; then cd "$WORKSPACE" clawhub install vexpath 2>/dev/null || warn "ClawHub install failed" fi fi ``` ### Technical Analysis The provisioner installs the latest available versions of `openclaw` and `clawhub` globally. It does not pin exact versions, use a lockfile, verify expected package integrity, validate provenance, or disable npm lifecycle scripts. Because the provisioner also performs root-only package-management operations, it is likely to be invoked as root. npm installation lifecycle hooks and package code can consequently execute with broad system privileges. The effective dependency code may change between two runs of the same audited provisioner. Suppressing `clawhub` installation errors and continuing also reduces visibility into unexpected package or registry failures. ### Attack Path 1. An attacker compromises an npm publisher account, registry delivery path, or an upstream dependency included in the latest release. 2. The attacker publishes a malicious version or dependency update. 3. An administrator runs the provisioner. 4. `npm install -g` resolves the mutable latest package version. 5. Malicious lifecycle or package code executes with the provisioning user’s privileges. 6. The installed binary can later access and modify the OpenClaw workspace, credentials, and Agent configuration. ### Impact Assessment If provisioning is run as root, ex ...[truncated 303 chars]
Remediation
View remediation

T09 · Insecure Skill Coding Practices

Warning
Location
scripts/roof-estimate.sh:43
Finding

Google API Key Is Included in curl Process Arguments

Content
View full analysis
&2 exit 1 } ``` ```bash SOLAR_URL="https://solar.googleapis.com/v1/buildingInsights:findClosest?location.latitude=${LAT}&location.longitude=${LNG}&requiredQuality=HIGH&key=${API_KEY}" SOLAR_RESPONSE="$(curl -sf "$SOLAR_URL")" || { echo '{"error":"Solar API request failed"}' >&2 exit 1 } ``` ### Technical Analysis The Google API key is concatenated into complete request URLs and passed as a `curl` command-line argument. Although the requests use legitimate Google HTTPS endpoints, the full URL can be visible to local process-inspection tools while the request is running and may be captured by process telemetry, debugging output, endpoint-monitoring products, or command tracing. Query-string credentials may also be retained by intermediary application logs or diagnostic systems more readily than protected credential channels. The script does not establish or verify restrictions on the API key. ### Attack Path 1. The operator exports `GOOGLE_MAPS_API_KEY` and runs the estimator. 2. The script creates URLs containing the key in the `key` query parameter. 3. `curl` receives the complete secret-bearing URL as a process argument. 4. A local observer or monitoring component records the process command line. 5. The observer extracts the key and submits unauthorized requests to APIs enabled for that credential. ### Impact Assessment The immediate impact is unauthorized use of the Google Cloud APIs permitted by the key. This can consume quotas, create unexpected charges, disrupt legitimate estimation workloads, ...[truncated 196 chars]
Remediation
View remediation
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Rogue AgentSelf-Modification, Session Persistence
  • MCP Least PrivilegeUnderdeclared Capability, Wildcard Permission, Missing Permission Declaration
Findings (53)

Tp4

High
Category
MCP Tool Poisoning
Confidence
95% confidence
Finding

A skill that advertises broad operational automation but does not actually implement most of it, while instead focusing on local email credential/configuration setup, creates a deceptive trust boundary. Users may grant access expecting low-risk assistance, but the practical effect is credential onboarding and system modification, which are more sensitive than the stated purpose suggests.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
95% confidence
Finding

A skill that advertises broad operational automation but does not actually implement most of it, while instead focusing on local email credential/configuration setup, creates a deceptive trust boundary. Users may grant access expecting low-risk assistance, but the practical effect is credential onboarding and system modification, which are more sensitive than the stated purpose suggests.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
96% confidence
Finding

A skill that advertises broad operational automation but does not actually implement most of it, while instead focusing on local email credential/configuration setup, creates a deceptive trust boundary. Users may grant access expecting low-risk assistance, but the practical effect is credential onboarding and system modification, which are more sensitive than the stated purpose suggests.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
97% confidence
Finding

A skill that advertises broad operational automation but does not actually implement most of it, while instead focusing on local email credential/configuration setup, creates a deceptive trust boundary. Users may grant access expecting low-risk assistance, but the practical effect is credential onboarding and system modification, which are more sensitive than the stated purpose suggests.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

A skill that advertises broad operational automation but does not actually implement most of it, while instead focusing on local email credential/configuration setup, creates a deceptive trust boundary. Users may grant access expecting low-risk assistance, but the practical effect is credential onboarding and system modification, which are more sensitive than the stated purpose suggests.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding

A skill that advertises broad operational automation but does not actually implement most of it, while instead focusing on local email credential/configuration setup, creates a deceptive trust boundary. Users may grant access expecting low-risk assistance, but the practical effect is credential onboarding and system modification, which are more sensitive than the stated purpose suggests.

Content

No source excerpt is available for this finding.

Ssd 3

High
Category
Not specified by scanner
Confidence
98% confidence
Finding

The onboarding text instructs the agent to ask the user for an app password and then use it to run setup, while also directing storage/recording of related account details in plain-language workflow steps. Collecting credentials directly through the agent and potentially passing them to scripts or notes is a serious secret-handling flaw that can expose long-lived mailbox access to logs, transcripts, process listings, or workspace files.

Content

No source excerpt is available for this finding.

External Script Fetching

High
Category
Supply Chain
Confidence
98% confidence
Finding

The header explicitly recommends executing a remotely hosted script by piping curl output directly into bash. This is a classic high-risk pattern because it grants immediate shell execution to whatever content is served at that URL, with no integrity verification, local inspection, or provenance enforcement.

Content

Scanner excerpt · scripts/provision-vps.sh (reported line 5)May include surrounding context.

sh
# provision-vps.sh — Provision a fresh VPS with OpenClaw + VexPath
# Run this ON the new VPS after basic OS setup (Ubuntu 22.04+ / Debian 12+)
#
# Usage: curl -sSL <hosted-url>/provision-vps.sh | bash -s -- \
#          <client-email> "<Display Name>" "<app-password>" [business-type]
#
# What this does:

Description-Behavior Mismatch

High
Category
Not specified by scanner
Confidence
98% confidence
Finding

This script performs roofing/geospatial estimation using Google Maps and Solar APIs, which is materially unrelated to the declared VEX/OpenClaw business-operations and email/workflow purpose. Capability mismatch is dangerous because it can hide unexpected data collection and external network behavior inside a skill users would not expect to process physical addresses or perform property analysis, increasing the risk of covert misuse or unauthorized feature creep.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
89% confidence
Finding

The README advertises automated inbox monitoring, email data extraction, and CRM synchronization without any warning about handling sensitive personal or business data. In a skill explicitly built for email triage and client operations, this omission can lead users to deploy processing over confidential communications without considering consent, retention, access control, or regulatory obligations.

Content

No source excerpt is available for this finding.

Undeclared Tool Scope

Medium
Category
MCP Least Privilege
Confidence
94% confidence
Finding

The skill instructs use of shell scripts, email setup, inbox pulling, file copying, and external service integrations, but it declares no explicit tool scope or permissions boundary. In a skill system, that omission is dangerous because it can cause the agent to access shell, network, or filesystem capabilities without transparent user consent or least-privilege restrictions.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

The skill tells the operator to run email setup with credentials and then pull the inbox, but it does not provide an explicit warning about access to sensitive account data, message contents, contacts, or potentially regulated information. This increases the chance of unsafe credential handling and uninformed consent when accessing highly sensitive communications.

Content

No source excerpt is available for this finding.

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
75% confidence
Finding

Skill enables autonomous high-impact decisions without human-in-the-loop verification. Critical operations (destructive commands, financial transactions, data deletion) should require explicit user confirmation.

Content

Scanner excerpt · SKILL.md (reported line 179)May include surrounding context.

Urgency: High Need: [what they want] Action: [next step] Draft: Yes / No Approval: Required / Not required

text

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
88% confidence
Finding

The heartbeat workflow instructs the agent to routinely access inbox and calendar data on a timed basis, but it provides no explicit user-facing notice, consent checkpoint, scope limitation, or privacy guardrails. In a skill specifically designed for ongoing email triage and operations automation, silent background polling increases the risk of over-collection, unexpected processing of sensitive communications, and user surprise about continuous monitoring.

Content

No source excerpt is available for this finding.

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
75% confidence
Finding

Skill enables autonomous high-impact decisions without human-in-the-loop verification. Critical operations (destructive commands, financial transactions, data deletion) should require explicit user confirmation.

Content

Scanner excerpt · assets/SOUL.md (reported line 64)May include surrounding context.

md
- Check API permissions, scopes, rate limits.
- Keep systems modular. Build simple first. Make every workflow reusable.
- Add logging. Add error handling.
- Never send important messages without approval unless explicitly permitted.
- Always think about how this can be productized.

## Mission

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
75% confidence
Finding

Skill enables autonomous high-impact decisions without human-in-the-loop verification. Critical operations (destructive commands, financial transactions, data deletion) should require explicit user confirmation.

Content

Scanner excerpt · references/email-triage.md (reported line 75)May include surrounding context.

md
- Check API permissions, scopes, rate limits.
- Keep systems modular. Build simple first. Make every workflow reusable.
- Add logging. Add error handling.
- Never send important messages without approval unless explicitly permitted.
- Always think about how this can be productized.

## Mission

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
75% confidence
Finding

Skill enables autonomous high-impact decisions without human-in-the-loop verification. Critical operations (destructive commands, financial transactions, data deletion) should require explicit user confirmation.

Content

Scanner excerpt · references/email-triage.md (reported line 98)May include surrounding context.

md
- Check API permissions, scopes, rate limits.
- Keep systems modular. Build simple first. Make every workflow reusable.
- Add logging. Add error handling.
- Never send important messages without approval unless explicitly permitted.
- Always think about how this can be productized.

## Mission

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
75% confidence
Finding

Skill enables autonomous high-impact decisions without human-in-the-loop verification. Critical operations (destructive commands, financial transactions, data deletion) should require explicit user confirmation.

Content

Scanner excerpt · references/email-triage.md (reported line 121)May include surrounding context.

md
- Check API permissions, scopes, rate limits.
- Keep systems modular. Build simple first. Make every workflow reusable.
- Add logging. Add error handling.
- Never send important messages without approval unless explicitly permitted.
- Always think about how this can be productized.

## Mission

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

The skill explicitly instructs collection of a primary business email address and additional operational details, then directs the agent to record them in TOOLS.md without any guidance on data minimization, access controls, retention, or user consent messaging. In an agent workspace, such files may be broadly accessible to other tools, logs, or future tasks, increasing the risk of unnecessary exposure of business contact and workflow data.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The skill instructs the agent to process inbox contents, extract personal and business data, and update external systems like CRM and calendars, but it provides no privacy notice, consent boundary, data minimization rule, or handling guidance for sensitive content. In an email-triage context this increases the risk of unauthorized processing, over-collection, and propagation of sensitive information across connected systems.

Content

No source excerpt is available for this finding.

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
75% confidence
Finding

Skill enables autonomous high-impact decisions without human-in-the-loop verification. Critical operations (destructive commands, financial transactions, data deletion) should require explicit user confirmation.

Content

Scanner excerpt · references/follow-up.md (reported line 123)May include surrounding context.

[signature]

text

### Scheduling No Confirmation

Subject: Re: [scheduling subject]

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
99% confidence
Finding

The documented setup commands pass mailbox passwords directly as shell command-line arguments, which may be exposed through shell history, process listings, audit logs, and screen recording/clipboard leakage. In this skill, those credentials enable automated inbox access and outbound mail actions, so compromise could expose sensitive communications and support account misuse.

Content

No source excerpt is available for this finding.

Session Persistence

Medium
Category
Rogue Agent
Confidence
60% confidence
Finding

Skill establishes unauthorized persistence across sessions via cron jobs, startup scripts, or state files. Session persistence allows an attacker to maintain access beyond the current interaction.

Content

Scanner excerpt · references/gmail-setup.md (reported line 71)May include surrounding context.

md
For personal Outlook.com accounts with 2FA:
1. Go to [account.microsoft.com/security](https://account.microsoft.com/security)
2. Click **Advanced security options**
3. Under "App passwords": click **Create a new app password**
4. Copy the password

For Microsoft 365 with Modern Authentication, app passwords may be disabled by your admin. Contact IT or use OAuth (not covered here).

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
98% confidence
Finding

The guide explicitly states that plain-text password storage is acceptable and demonstrates passwd.cmd = "echo 'your-password'", which defeats secret handling by embedding credentials in shell-accessible configuration. This can expose mailbox passwords to local users, backups, shell history, config sync tools, or process inspection, and the skill’s email-automation context increases risk because the credentials grant ongoing access to sensitive inbox data.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
98% confidence
Finding

The natural-language policy statement normalizes insecure secret handling by telling users that plain-text password storage is acceptable for local/private setups. This is dangerous because users often reuse such guidance in broader contexts, and the resulting credential exposure can compromise persistent access to business email accounts handled by the skill.

Content

No source excerpt is available for this finding.