T01 · Skill Instruction Hijacking
- Location
SKILL.en.md:8- Finding
Persistent Final-Response Manipulation Through Mandatory Signatures
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
The skill is a coherent sign-off customizer with no evidence of malicious code, though it can persistently change response formatting.
Install this only if you want the assistant to append a signature to ordinary completed responses. Disable or override it for strict JSON, code-only, checksum, command-output, or automation workflows, and review any requested changes to sign-off.json before letting them persist.
SKILL.en.md:8Persistent Final-Response Manipulation Through Mandatory Signatures
The skill’s stated purpose is a simple sign-off marker, but it also instructs the agent to obtain weather data via an external API. That expands scope from formatting output into network-enabled enrichment, which can introduce unnecessary data exposure, hidden dependencies, and surprising side effects for users who did not consent to external calls.
The documentation allows conversational commands to update sign-off.json, giving the skill persistent file-write capability beyond merely appending a signature. Persistent state changes from natural-language requests are risky because they can be triggered unintentionally or through prompt injection, and they are not clearly framed as a write operation.
The skill indicates that user conversation can update sign-off.json but does not warn that this modifies a workspace file. Undisclosed persistence is dangerous because users may believe they are making a temporary stylistic request when they are actually changing future agent behavior on disk.
Template installation modifies workspace files without disclosing that side effect to the user. Hidden writes are especially risky in agent skills because they create persistent behavioral changes that can outlast the immediate session and may be abused through indirect instructions.
Without declared permissions the skill's intent is opaque and cannot be validated.
The description mixes Chinese instructions and directs the agent to append a sign-off to every response, but does not offer users any language or locale choice for the sign-off behavior. Because the skill content is written as a fixed Chinese-language interaction pattern rather than an opt-in locale choice, it may conflict with language/locale policy expectations.
The skill tells the agent to update sign-off.json based on a user's natural-language request, but it does not require explicit disclosure or confirmation that a workspace file will be modified. This can cause silent persistent state changes, surprising users and enabling prompt-driven configuration tampering across later interactions.
The skill instructs installing templates by reading from templates/ and merging data into sign-off.json without telling the user that a persistent file change will occur. Hidden config mutation increases the risk of unreviewed state changes and can be abused to alter future agent behavior through seemingly harmless style requests.
The entire skill documentation is written in Chinese and presents the variables, modes, and output examples exclusively in Chinese, implying a fixed language/locale expectation. Under the policy rule, forcing a specific language without user opt-in can be a natural-language policy violation unless the locale constraint is explicitly justified, which is not present here.
The package description explicitly states the skill adds a personalized sign-off in Chinese-style wording without indicating any user choice, locale detection, or opt-in. A skill that unconditionally alters response language or format can override user expectations, reduce usability for non-Chinese speakers, and create prompt-injection-like policy conflicts when system behavior is supposed to remain neutral or user-directed.
The file includes its description, comments, greeting strings, and day/time labels entirely in Chinese, which effectively forces a specific language in the generated output. The policy requires user choice or a clearly documented locale-specific justification, neither of which appears in this file.
The description and template specify a Japanese haiku style using Japanese text and particles (for example, "の" and the Japanese preview), with no indication that users can choose another language or locale. This creates a language/locale policy issue because the skill appears to enforce a specific language presentation by default.
The display name, description, and preview are all written in Chinese, which indicates the skill is oriented to a specific language/locale without any visible opt-in, alternative locale support, or justification that it is region-specific. This matches the policy category for language or locale constraints expressed in natural-language content.
This JSON template embeds Chinese natural-language strings in the display name, description, and preview, which indicates the skill is designed to present itself and operate in a specific language/locale by default. The file does not offer any user choice or opt-in for language, so it conflicts with the policy against forcing a specific language without user opt-in.
The file includes user-facing natural-language examples in both Chinese and English, but one example directly tells the AI how to change its signature using Chinese phrasing. In the absence of an explicit statement that language should follow user preference, this may create a mild locale-policy concern if the skill nudges the assistant toward a specific language style.
Reading from a templates/ library and merging data into sign-off.json broadens the skill into template management and file mutation rather than simple output decoration. While lower severity than arbitrary file access, it still increases attack surface through untrusted template content, unsafe merges, or unexpected persistent configuration changes.
The display name, description, and template are written to enforce a Chinese 'ancient style' output format, with no indication that users can choose another language or locale. The policy allows locale constraints when they are clearly documented and justified, but this file does not state a regional-only purpose or provide opt-in.
The display name and description specify a Chinese-language style ("极客风", "程序员风格,简洁精确") with no indication that users can choose another language or locale. This can violate language/locale policy when a skill imposes a specific language presentation by default rather than offering opt-in or selection.
This manifest hard-codes both Chinese-language metadata and English-rendered template output, but provides no user opt-in or locale selection. That can violate language/locale policy expectations when users or deployments require language choice rather than a fixed mixed-locale presentation.
The display name, description, and preview are written only in Chinese, which indicates the skill is designed to present output in a specific language without offering any user choice or opt-in. The policy for natural-language violations applies to all file types and includes locale or language constraints that are imposed implicitly.