Back to skill

Security audit

Hormuzmonitorskills

Security checks for vulnerabilities and agentic risk

Overview

The skill’s monitoring goal is coherent, but it tells an agent to run an unreviewed local deployment script that can push to GitHub and trigger Vercel without a clear approval step.

Install only if you control and trust the referenced ~/hormuz-website/scripts/update_and_deploy.sh script and are comfortable with the agent modifying MONITOR_LOG.md, deleting entries older than 7 days, and pushing website data to GitHub/Vercel. Prefer adding a manual review and approval step before any push or deployment.

Vulnerability Patterns
  • Tool Hijacking and SpoofingModifies or replaces tools so legitimate-looking calls execute attacker logic
  • Skill Instruction HijackingAlters the agent's session goals or safety constraints when the skill loads
  • Agent Memory PoisoningWrites attacker-controlled rules into memory that affect later sessions
  • Remote Payload Retrieval and ExecutionFetches external code whose behavior can change after review
  • Embedded Malicious CodeShips malicious scripts inside the skill and executes them locally
Findings (1)

T07 · Tool Hijacking and Spoofing

Error
Location
SKILL.md:118
Finding

Execution of an Unverified Mutable Deployment Script

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, lines 118-125
Vulnerability Type: Unverified external local script execution
Risk Level: High

Vulnerable Code

markdown
### Step 7: Update Website Data
After writing MONITOR_LOG.md, run this command to update the website:

exec command: bash ~/hormuz-website/scripts/update_and_deploy.sh

This parses MONITOR_LOG.md into data.json and pushes to GitHub.
Vercel auto-deploys on push. If the script fails, log the error but
do not retry — the website will show the previous data until next cycle.

Technical Analysis

The Skill instructs the agent to invoke bash on a script located outside the reviewed package. The referenced ~/hormuz-website/scripts/update_and_deploy.sh file was not included in the project and therefore could not be inspected. No integrity hash, trusted ownership check, permission validation, canonical path validation, or content verification is required before execution.

Because the script is resolved beneath the runtime user's home directory, any party or process capable of creating or modifying that path can control the commands executed during Step 7. Explicitly invoking the file through bash executes its contents regardless of whether the file itself has executable permissions.

The documented purpose of the script includes parsing data, pushing changes to GitHub, and initiating a Vercel deployment. It may consequently operate in an environment containing repository or deployment credentials, increasing the impact of script substitution.

Attack Path

  1. An attacker gains write access to the runtime user's ~/hormuz-website directory, the scripts subdirectory, or update_and_deploy.sh.
  2. The attacker creates or replaces update_and_deploy.sh with commands that perform unauthorized actions, such as reading accessible credentials, modifying repository content, or executing additional local commands.
  3. The monitoring workflow completes its collection a ...[truncated 1315 chars]
Remediation
View remediation

Remediation Suggestions

  1. Include the deployment script in the reviewed Skill package so its exact contents and changes can be audited.
  2. Reference the script through a fixed, canonical absolute path rather than a home-directory-relative path.
  3. Before execution, verify that the script and every parent directory are owned by the expected trusted account and are not writable by untrusted users or groups.
  4. Pin and validate a cryptographic hash or signed release of the script before invoking it. Abort deployment when verification fails.
  5. Execute the deployment process under a dedicated least-privileged account with access only to the required repository and deployment target.
  6. Use narrowly scoped, short-lived GitHub and deployment credentials. Do not expose unrelated credentials to the script environment.
  7. Require explicit operator approval before repository pushes or production deployment, particularly when the script or generated changes differ from expected content.
  8. Restrict the script's environment, filesystem access, and network destinations through sandboxing or container isolation.
  9. Review and validate the generated repository diff before pushing. Permit changes only to expected files such as the intended data.json.
  10. Log script identity, verified hash, generated changes, deployment result, and failures without recording secrets.
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
Findings (6)

Context-Inappropriate Capability

High
Category
Not specified by scanner
Confidence
98% confidence
Finding

The skill explicitly instructs execution of a local shell script that updates data and pushes changes to GitHub, which creates a direct bridge from untrusted fetched content to local code execution and external deployment. Even if the script path is fixed, this is dangerous because it performs side effects on the host and a remote repository without an explicit approval gate or validation boundary.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The manifest description says the skill will monitor shipping traffic, write findings, and auto-update website data, but it does not specify the exact invocation conditions or limiting context for when this skill should run. In a manifest file, this broad description can create ambiguous activation scope because there are no explicit trigger phrases, exclusions, or negative examples.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
98% confidence
Finding

The instructions require all summary fields to be written in Chinese, which forces a specific language regardless of user preference. This is a natural-language policy issue because no user choice, opt-in, or documented locale justification is provided.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The skill directs the agent to delete historical entries older than 7 days from MONITOR_LOG.md without warning or confirmation. Silent data removal can destroy audit history, impair incident reconstruction, and cause integrity issues if the log is relied on for monitoring or downstream automation.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

These instructions combine subprocess execution with deployment to external services but provide no user warning or consent checkpoint. In context, the skill is not merely formatting text; it is telling the agent to perform consequential actions on the local machine and publish data externally, which materially increases the risk of unauthorized changes or exfiltration.

Content

No source excerpt is available for this finding.

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
95% confidence
Finding

The skill endorses autonomous decision making that culminates in automatic website deployment after each cycle. This is dangerous because the agent is empowered to make publication decisions based on scraped third-party content and then trigger remote state changes without a human verification step, increasing the chance of bad data propagation or unintended operational impact.

Content

Scanner excerpt · SKILL.md (reported line 124)May include surrounding context.

md
exec command: bash ~/hormuz-website/scripts/update_and_deploy.sh

This parses MONITOR_LOG.md into data.json and pushes to GitHub.
Vercel auto-deploys on push. If the script fails, log the error but
do not retry — the website will show the previous data until next cycle.

## Alert format