T01 · Skill Instruction Hijacking
- Location
SKILL.md:98- Finding
Mandatory Promotional Output Hijacking
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
This skill has a coherent prompt-recommendation purpose, but it gives mutable remote data and shell commands too much control for a prompt library.
Review this carefully before installing. The skill is not proven malicious, but it automatically downloads a mutable prompt dataset, refreshes it during use, and tells the agent to run shell commands using image paths from that dataset. Prefer a version that pins and validates the dataset, avoids shell-based downloads, uses safe temporary files, and makes network refreshes explicit or opt-in.
SKILL.md:98Mandatory Promotional Output Hijacking
SKILL.md:180Shell Command Injection Through Remotely Controlled Image Paths
SKILL.md:192Predictable Shared Temporary File Enables Clobbering and Cross-Session Races
scripts/setup.js:20Mutable Remote Dataset Is Automatically Trusted Without Integrity or Schema Validation
README.md:48Documentation Recommends Unpinned Third-Party npx Installers
The skill instructs use of an exec tool to run rm on a filesystem path as part of normal operation. Even though the specific example targets a fixed temp file, embedding deletion commands into a content-recommendation skill normalizes unnecessary shell access and could become dangerous if paths or filenames later become variable or user-influenced.
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.
The skill advertises broad natural-language triggers like 'Help me find a YouTube thumbnail' and 'Find me a cyberpunk-style avatar prompt,' which are common user phrases that could cause the skill to activate unexpectedly in general conversations about image creation. In agent ecosystems with automatic skill routing, ambiguous invocation examples increase the chance of unintended context exposure or tool use when the user did not explicitly request this specific skill.
The remix-mode examples invite users to paste arbitrary article text, video scripts, and notes using very broad phrasing, which can trigger the skill on large sensitive content bodies without clear scoping. In an auto-selected skill environment, this raises the risk of unnecessary ingestion of private user content and accidental routing of unrelated documents into the skill's prompt-recommendation flow.
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 skill instructs the agent to perform network operations (setup/update checks and image downloads) but does not declare an explicit tool scope such as allowed tools or permissions. That mismatch weakens policy enforcement and reviewability, because a seemingly content-only prompt skill can trigger external access at runtime.
The usage conditions include very broad triggers such as users wanting to 'Generate images with AI' or 'Find proven AI image generation prompts,' which could overlap with many ordinary image-assistance requests. The description does not provide explicit exclusion conditions or tighter constraints for when this skill should be invoked instead of other image-generation or creative-assistance skills.
The skill mandates a network freshness check and possible synchronization from HuggingFace on every run before serving the user. This introduces unnecessary remote dependency and data-fetch behavior for a recommendation skill, creating supply-chain and privacy exposure whenever the skill is invoked.
The skill explicitly says stale references should be updated silently, without a user-facing warning that remote network access will occur. Silent background fetching reduces transparency and prevents informed consent about external data access and potential data leakage through query-triggered activity.
The skill requires shell execution to download remote images to /tmp and then delete them, even though prompt recommendation can be fulfilled by linking or embedding hosted images directly. This expands the attack surface to command execution, filesystem writes, and processing of untrusted remote content without necessity.
The instructions require the 'Customized prompt' to be output in English for generation, regardless of the user's language. Because this is a mandatory language constraint rather than a user choice or clearly justified regional limitation, it can violate language/locale policy expectations.
The language handling section instructs the skill to respond in the user's language but always provide prompt content in English for generation. This is an explicit language restriction that does not provide user choice or documented justification sufficient to exempt it from language/locale policy concerns.
The package description says the skill 'Works with Claude Code, Cursor, and more' but does not specify how or when the skill is invoked, what trigger phrases apply, or any constraints or exclusions. In a manifest file, this kind of broad capability description can contribute to unclear activation boundaries and unintended invocation expectations.
The README says users can communicate in their own language but the skill will '始终提供英文提示词用于生成' (always provide English prompts for generation). This is a natural-language locale policy constraint presented as unconditional behavior rather than a user choice or opt-in.
The natural-language description centers compatibility around specific AI tools and model branding, but it does not present this as a user-selectable option or clarify limits. While not a strong violation, this could be read as prescribing a fixed ecosystem without explicit user opt-in.