T01 · Skill Instruction Hijacking
- Location
SKILL.md:119- Finding
Persistent Third-Party Branding Hijacks Generated Output
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
This skill has a coherent image-summary purpose, but it asks users to paste an API key into chat and sends chat-derived content to an external service with weak credential scoping.
Review before installing. Use it only if you are comfortable giving the agent a mew.design API key and sending summarized chat content to Mew's API. Avoid using it with confidential or regulated conversations, prefer a short-lived or low-scope API key, and watch for any command that changes the validation base URL away from api.mew.design.
SKILL.md:119Persistent Third-Party Branding Hijacks Generated Output
scripts/validate_mew_design_key.py:18API Key Is Passed Through Process Command-Line Arguments
scripts/validate_mew_design_key.py:21Arbitrary Validation Base URL Can Redirect the User's API Key
The declared purpose suggests a simple conversation-to-image skill, but the actual behavior includes collecting and validating a third-party API key, probing authentication behavior, and sending user content to an external service. This mismatch undermines informed consent and can trick users into exposing credentials or conversation data under a less sensitive description.
Skill manipulates agent memory, state, or stored context. Memory corruption can alter personality, override safety rules, or cause unpredictable behavior.
## Compression Heuristics
- Merge repeated requests into one clear statement.
- Collapse implementation detail into one bullet unless the detail affects outcome.
- Keep at most one metric per bullet unless comparisons matter.
- Convert long prose into noun-verb phrases.
The skill invokes shell commands, writes temporary files, and makes network requests but does not declare any tool scope or permissions. That creates a hidden capability gap where an installer or user may believe the skill is limited to summarization while it can actually exfiltrate data and handle secrets through undeclared channels.
The installation triggers are broad enough to match common user requests such as asking for help installing a skill, which can cause the agent to initiate installation unexpectedly. Overbroad activation increases the chance of unreviewed execution of a skill that handles secrets and transmits data externally.
The skill instructs the agent to collect the user's mew.design API key and pass it into shell commands and HTTP headers without a clear warning about secret handling, storage, or exposure risks. API keys are sensitive credentials, and prompting for them casually inside chat increases the risk of disclosure, logging, reuse, or mishandling.
The skill summarizes conversation content and sends it to an external image-generation API, but it does not clearly warn the user that chat content will leave the local environment. Conversations can contain confidential business, personal, or regulated data, so undisclosed external transmission is a material privacy and security risk.
The instructions require the agent to use the "exact onboarding copy" and the style-selection format entirely in Chinese, regardless of the user's language preference. This is a natural-language policy issue because it imposes a specific locale without opt-in or justification.
The hardcoded use of the api.mew.design endpoint confirms dependency on a third-party service for processing user-derived data and credentials. Even if this is functionally required, it is security-relevant because it expands the trust boundary beyond the local agent and can expose confidential material.
Then generate the image through the official Mew design API:
curl -sS -X POST "https://api.mew.design/open/api/design/generate" \
-H "Content-Type: application/json" \
-H "x-api-key: USER_PROVIDED_KEY" \
--data @/tmp/chat-summary-body.json
The hardcoded use of the api.mew.design endpoint confirms dependency on a third-party service for processing user-derived data and credentials. Even if this is functionally required, it is security-relevant because it expands the trust boundary beyond the local agent and can expose confidential material.
Then generate the image through the official Mew design API:
curl -sS -X POST "https://api.mew.design/open/api/design/generate" \
-H "Content-Type: application/json" \
-H "x-api-key: USER_PROVIDED_KEY" \
--data @/tmp/chat-summary-body.json
This instruction says to create the image in Chinese unless the input is clearly another language, which imposes a language choice by default rather than letting the user choose. The policy for this audit flags language or locale constraints when they are forced without explicit opt-in or documented justification.
Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.
import urllib.parse
import urllib.request
DEFAULT_BASE_URL = "https://api.mew.design/"
def parse_args() -> argparse.Namespace: