T01 · Skill Instruction Hijacking
- Location
SKILL.md:22- Finding
Broad Agent Intent and Execution Hijacking
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
This load-testing skill is purpose-aligned but needs review because it can automatically create live load-test tasks for broad website requests using a shared token and unsafe TLS settings.
Install only if you intend to run live LMeterX load tests on systems you own or are explicitly authorized to test. Require manual confirmation before task creation, replace the shared token with a user-specific secret, avoid broad automatic URL triggers, and fix TLS certificate verification before using this in a real environment.
SKILL.md:22Broad Agent Intent and Execution Hijacking
scripts/run.py:475TLS Certificate Verification Disabled for Authenticated Backend Requests
Hidden instructions were detected in comments or invisible text. These could contain malicious directives. Manual review is recommended.
### 1. Environment Variable Configuration
The script comes with the following built-in default values and can be run directly. If you wish to override these defaults, you can set variables with the same names via environment variables or the `.openclaw/.env` file:
| Variable | Default Value | Description |
|------|--------|------|
The skill instructs immediate execution of load-testing actions against any user-provided URL with no health checks, authorization validation, or warning about disruptive effects. Load testing can degrade service availability or resemble denial-of-service behavior, so automatic execution materially increases the risk of abuse or accidental harm to third-party systems.
The skill mandates invocation whenever a user mentions general web-related terms or provides any website-like URL, creating a strong risk of accidental activation. In this context, accidental activation is especially dangerous because the skill immediately initiates analysis and load-test task creation against external targets, which can cause unauthorized or disruptive activity.
The README explicitly promotes automatic page analysis, API discovery, and batch creation of load-testing tasks from a supplied URL, but it does not warn that users must have authorization from the target owner or that active load generation can disrupt services. In the context of an agent skill that triggers directly from natural-language requests, this omission materially increases the risk of misuse for unauthorized scanning or denial-of-service activity.
The skill invokes a Python script that uses environment variables and performs networked actions, yet the manifest declares no explicit tool scope or permission boundaries. In a load-testing skill, this omission is risky because it obscures that the skill can make outbound requests and use credentials, reducing reviewability and increasing the chance of unsafe execution in unintended contexts.
The listed trigger phrases are broad, natural-language requests that overlap with ordinary conversation and can cause the skill to run when the user did not clearly intend a real load test. Because the skill is capable of immediate outbound actions, this broad matching increases the likelihood of unintended testing against live systems.
The documentation hard-codes and exposes a built-in authentication token and instructs operators to export it for use, which normalizes credential reuse and weakens access control. Even if the token is low-privilege, embedding credentials in skill instructions encourages unauthorized use of the LMeterX service and makes misuse easier to scale.
The script globally disables TLS certificate verification with httpx.Client(verify=False), which allows a man-in-the-middle attacker to intercept or modify all HTTPS traffic to the LMeterX backend. In this skill, those requests carry the X-Authorization token and control load-test task creation, so an attacker could steal credentials, alter analyzed API configs, or redirect task creation results.
verify=False is an unsafe default because it disables server certificate validation for every outbound HTTPS request. Given this skill immediately analyzes user-supplied URLs and then sends authenticated backend requests that create load-testing tasks, the lack of TLS verification materially increases the risk of token theft, response tampering, and malicious task manipulation if network traffic is intercepted.
_validate_url(args.url)
_validate_concurrency(args.concurrent_users)
with httpx.Client(verify=False) as client:
# Step 1: Analyze page
print(f"\n🔍 Step 1/3: Analyze page {args.url} ...")
configs, summary = analyze_page(
The manifest hard-codes activation phrases in English and Chinese, but does not indicate whether users may interact in other languages or whether this bilingual restriction is intentional. This can create an implicit language-policy constraint without opt-in or explanation.
The instructions say the agent is 'strictly forbidden' from constructing HTTP requests manually and may only execute a local script, but later documentation states that the script automatically performs API connectivity pre-checks and load-testing task creation, which necessarily involve HTTP/API calls. This is an intent/documentation contradiction about whether network API interaction is part of the skill's behavior versus prohibited.
The timeout error string includes Chinese text ('请缩小分析范围') in an otherwise English-language script output. This imposes a specific language on users without any opt-in or documented locale constraint, which matches the natural-language locale policy concern.