T01 · Skill Instruction Hijacking
- Location
SKILL.md:100- Finding
Mandatory Promotional Output Injection
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
This skill does what it advertises, but it automatically downloads mutable remote data and tells the agent to run shell commands using fields from that data, which creates a real review-worthy safety concern.
Review before installing. The prompt library and sample images are public, but installation and normal use can contact HuggingFace and update local files. The main risk is the skill's instruction to run shell curl commands using mutable remote metadata; only install it in an environment where command execution is constrained, and avoid pasting sensitive content into the remix workflow unless you are comfortable with how your agent processes it.
SKILL.md:100Mandatory Promotional Output Injection
SKILL.md:180Shell Command Injection Through Remotely Controlled Image Metadata
The instruction rm /tmp/prompt_img.jpg is a concrete command-execution pattern that normalizes destructive shell use in a skill whose purpose is only prompt recommendation. Even though the shown filename is fixed, this pattern is dangerous because it encourages use of unrestricted exec for file operations; if the path ever becomes variable or attacker-influenced, it could lead to deletion of unintended files or command-injection style abuse.
message tool: action=send, media=/tmp/prompt_img.jpg, caption="[Prompt Title]"
Step C — Cleanup: exec: rm /tmp/prompt_img.jpg
Do this for **each** of the 3 recommended prompts — one image per prompt.
npx commands without a version suffix (e.g. @1.0.0) create a rug-pull risk if the upstream server is compromised and publishes a malicious update.
npx commands without a version suffix (e.g. @1.0.0) create a rug-pull risk if the upstream server is compromised and publishes a malicious update.
npx commands without a version suffix (e.g. @1.0.0) create a rug-pull risk if the upstream server is compromised and publishes a malicious update.
npx commands without a version suffix (e.g. @1.0.0) create a rug-pull risk if the upstream server is compromised and publishes a malicious update.
npx commands without a version suffix (e.g. @1.0.0) create a rug-pull risk if the upstream server is compromised and publishes a malicious update.
npx commands without a version suffix (e.g. @1.0.0) create a rug-pull risk if the upstream server is compromised and publishes a malicious update.
npx commands without a version suffix (e.g. @1.0.0) create a rug-pull risk if the upstream server is compromised and publishes a malicious update.
npx commands without a version suffix (e.g. @1.0.0) create a rug-pull risk if the upstream server is compromised and publishes a malicious update.
The README explicitly encourages users to paste full articles, video scripts, podcast notes, and similar content into the skill, but it gives no warning about sensitive, proprietary, or personal data handling. In an agent ecosystem where content may be processed by external tools, logged, or shared across model providers, this can lead to unintentional data disclosure.
The skill performs network-capable actions such as postinstall fetching, setup.js --check, and remote image downloads, but the manifest declares no tool scope or allowed-tools constraints. This creates a mismatch between what the skill appears to require and what it actually instructs the agent to do, reducing transparency and making unintended network access harder to audit or sandbox.
The skill mandates an automatic freshness check and silent network update on every run without clear user-facing consent. Repeated background fetching from a remote source expands supply-chain and privacy risk, and can cause the agent to perform external network activity users did not request or expect.
The skill instructs the agent to execute shell commands to download remote images and delete local files for each recommendation, which goes beyond normal prompt browsing. Unnecessary command execution increases attack surface, especially when file paths or URLs could later become influenced by untrusted dataset content or implementation changes.
The manifest description emphasizes recommending suitable prompts from a curated library and browsing sample prompts. In contrast, the workflow explicitly expands the skill into prompt creation and transformation, including AI-generated prompts and personalized remixes, which goes beyond straightforward recommendation from the 15,600+ dataset.
The package presents itself as a prompt library/recommendation skill, but its manifest also declares automatic setup code execution and auto-download behavior during installation. That mismatch expands the trust boundary: users expecting static prompt content may unknowingly run code and trigger network activity, which creates supply-chain and remote-content risks if the setup script or downloaded content is compromised.
A postinstall script causes code to run automatically whenever the package is installed, even though the stated purpose is only to provide image-generation prompts. In the context of a prompt-library skill, automatic execution is not necessary for core functionality and increases the chance of unnoticed malicious actions such as unexpected downloads, environment inspection, or persistence if the script is altered.
npx commands without a version suffix (e.g. @1.0.0) create a rug-pull risk if the upstream server is compromised and publishes a malicious update.
The README says the skill supports multilingual interaction but 'always provides English prompts for generation.' This is a language-policy constraint presented as mandatory behavior rather than a user choice, which fits the locale/language policy violation category.
The skill requires transient writes to /tmp and file deletion as part of ordinary recommendations, but this behavior is not surfaced in the skill description. While the immediate action is limited, undisclosed filesystem use reduces transparency and can become risky in constrained or multi-tenant environments.
The manifest description says the skill 'Works with Claude Code, Cursor, and more,' but does not specify the contexts, commands, or trigger boundaries under which it should be invoked. In a manifest file, this kind of broad activation framing can contribute to ambiguous triggering or unclear scope.