T09 · Insecure Skill Coding Practices
- Location
SKILL.md:323- Finding
Automatic Persistent Collection of User Profile Data Without Defined Privacy Controls
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
This coaching skill is not overtly malicious, but it needs Review because its documentation points to a different package/personality type and it directs automatic retention of personal coaching records without clear privacy controls.
Review carefully before installing. The package should align its README, manifest, install command, references, and versioning, and it should make growth-record storage opt-in with clear retention, deletion, and review controls. Users should avoid sharing sensitive personal, health, credential, financial, or employment details unless they know where the data is stored and how to delete it.
SKILL.md:323Automatic Persistent Collection of User Profile Data Without Defined Privacy Controls
README.md:1Documented Installation Command Resolves a Different Skill Package
The skill metadata declares an ISTJ-only coaching skill, but the README consistently documents an INTP coach with different triggering phrases, scope, and guidance. This mismatch can cause the wrong skill to activate for users, leading to misleading behavior, broken safety expectations, and accidental collection or processing of conversations under false pretenses.
Presenting the skill as 'INTP Coach' in documentation while the manifest declares an ISTJ skill is a direct intent contradiction. In an agent ecosystem, this can misroute users, undermine trust in skill selection, and create opportunities for deceptive or unintended invocation because operators and users cannot reliably tell what the skill will do.
The referenced file is clearly scoped to INTP coaching while the skill metadata declares an ISTJ coaching skill. This scope mismatch can cause the agent to deliver personality guidance to the wrong user segment, producing misleading or inappropriate advice and weakening trust in the skill's behavior.
The document is labeled as ISTJ guidance but the listed cognitive stack (Si→Te→Fi→Ne) and the repeated discussion of Ti-Ne loops, weak Fe, and broad-possibility traits are internally inconsistent and map to a different personality profile. In a coaching skill, this can systematically give users incorrect self-models and poor career/growth advice, especially because the file is intended as internal reference material that may shape all downstream responses.
The README states that growth records and user profiles are saved automatically, but it does not explain what data is stored, how long it is retained, where it is kept, or whether users can opt out. In a coaching context, conversations may include sensitive mental-health, career, and personal information, so silent retention increases privacy and consent risk.
The trigger description and all response templates are specified in Chinese, and the guidance assumes Chinese-language operation throughout. There is no indication that users may choose another language or that the Chinese-only constraint is intentional and justified for a region-specific use case.
The manifest explicitly states that vague emotional expressions such as '心情不好' without ISTJ context should not trigger the skill. In contrast, '场景 1' provides a handling flow and response template for exactly those inputs, including introducing itself and offering ISTJ coaching options, which is behavior beyond the stated trigger boundary.
The manifest description limits use to cases where the user explicitly says they are ISTJ and says non-ISTJ users should not trigger the skill. However, the skill instructions for '场景 2' explicitly tell the agent to continue offering the skill to users who say they are not ISTJ, which broadens the behavior beyond the declared scope.
The skill states it will save conversation and action records but provides no user-facing notice, consent flow, retention policy, or limits on what is stored. Because the skill handles personal growth, career, and emotional-context discussions, silent retention can collect sensitive behavioral data users may not expect to persist.
Retaining and reusing user-provided personal details across sessions creates persistent memory of potentially sensitive behavioral and emotional information. In a coaching context, such accumulation increases the blast radius of any leakage, misuse, or unintended recall because the data becomes longitudinal and identifiable.
Repeated instructions to log user issues, interests, actions, and insights amount to structured profiling over time. This is dangerous because it normalizes broad collection of personal data without demonstrating necessity, user awareness, or safeguards appropriate for reflective coaching conversations.
These instructions require recording profile-like personal data without a clear notice or consent checkpoint. Logging a user's problems, interests, actions, and insights creates a longitudinal personal dataset that could expose sensitive information if retained or later surfaced unexpectedly.
The growth-record section directs collection of detailed personal behavioral and conversational data for longitudinal tracking, which can reveal patterns about mental state, work habits, and decision-making. In a coaching skill, that profile can become highly sensitive even if each individual note seems harmless.
A mandatory instruction to maintain a user archive for future recall creates privacy risk by institutionalizing persistent storage as the default operating mode. The danger is amplified by the personal nature of the skill, where archived context may include intimate struggles, career uncertainty, and behavior patterns.
The end-of-conversation workflow mandates archiving key user information automatically, with no explicit user warning or approval. This is risky because users may believe they are having an ephemeral coaching conversation while the system is actually building a persistent record of personal issues, plans, and outcomes.
The end-of-conversation archiving instructions require storing user problems, insights, plans, and outcomes for later use, creating a persistent dossier of sensitive self-improvement information. If disclosed, misapplied, or surfaced in the wrong context, this could violate user expectations and expose private personal history.
The manifest metadata is materially inconsistent with the stated purpose of the skill: it is named and described as an ISTJ coach, but the keywords and tags prominently reference INTP instead. This kind of identity mismatch can cause the wrong skill to be surfaced or trusted under false pretenses, increasing the risk of misrouting users to incorrect guidance or hiding a repurposed package.
The file explicitly labels itself as 'INTP Coach' reference data near the end, directly contradicting the skill's ISTJ identity. This increases the likelihood that the skill will systematically condition responses using the wrong behavioral model, creating persistent misclassification and unsafe coaching recommendations for users relying on tailored guidance.
The entire reference file is written in Chinese and explicitly instructs that it is used internally by the skill during execution. For a coaching skill, embedding all operational guidance in a single language can create a language-policy issue if the skill behavior is effectively constrained to Chinese without offering the user a language choice or documenting a justified locale restriction.
The title and description frame the skill as 'INTP Coach' and the usage examples assume '我是 INTP', effectively steering the interaction toward a fixed user category. While not a locale issue, this is a natural-language constraint on who the skill is for without an explicit opt-in or broader applicability guidance.
The repository URL appears unrelated to the declared coaching skill, which undermines provenance and makes it harder to verify the source, maintenance history, and integrity of the package. In a skill ecosystem, mismatched repository links can conceal repackaged or misleading content and reduce reviewer ability to audit the artifact.
This markdown file is natural-language reference material, and its content consistently assumes Chinese as the required interaction language. Under the policy, forcing a specific language without offering a user choice or documenting a justified locale constraint is a natural-language policy concern.