T01 · Skill Instruction Hijacking
- Location
SKILL.md:661- Finding
Mandatory Promotional Content Injected into User-Facing Results
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
This skill is not clearly malicious, but it needs review because it handles sensitive mental-health material and includes unsafe local file commands that could delete or overwrite more than intended.
Review carefully before installing. Use only fictional, authorized, or strongly de-identified counselor material, do not rely on it for professional care or crisis support, and avoid running delete or rollback commands with untrusted counselor names until slug and path validation are added.
SKILL.md:661Mandatory Promotional Content Injected into User-Facing Results
SKILL.md:585Shell Command Injection and Arbitrary Deletion Through an Unvalidated Counselor Slug
scripts/version_manager.py:19Path Traversal Permits File Creation and Overwrite Outside the Version Store
The trigger phrases include broad emotional statements such as feeling depressed, stressed, or in a bad mood, which are likely to occur in ordinary conversation and can unintentionally invoke the skill. In the context of a psychology-themed skill, accidental activation is especially risky because users may disclose sensitive mental health information without realizing they are entering a specialized workflow.
The README encourages users to upload or paste extremely sensitive materials, including counseling records, case conceptualizations, supervision records, and social content, but does not provide a clear privacy notice, consent requirements, retention policy, or handling restrictions. In a mental health context, this is especially dangerous because the materials may contain protected health information, third-party disclosures, and confidential therapeutic content whose misuse or exposure could cause severe privacy harm.
These trigger phrases are broad emotional-distress statements such as feeling depressed, anxious, stressed, or wanting someone to talk to. That means the skill may auto-activate during real mental health or crisis-adjacent conversations and steer users into an unlicensed counselor simulation, creating a substantial risk of inappropriate reliance or delayed professional care.
The skill instructs execution of 'rm -rf counselors/{slug}' with a user-controlled parameter. If slug is not strictly validated, an attacker could inject path traversal, option-like values, or unexpected shell tokens, leading to deletion outside the intended directory and destructive data loss.
/delete-counselor {slug}: 确认后执行:
rm -rf counselors/{slug}
The English distress triggers are highly generic and overlap with ordinary help-seeking and potentially crisis-related statements. In a mental-health-themed skill, such broad matching is especially dangerous because it can insert a simulated therapist into vulnerable situations where stronger safety controls and escalation paths are needed.
Skill instructs the agent to omit warnings, disclaimers, or ethical commentary. Stripping safety caveats hides risk from the user and is a common jailbreak preamble.
### Listening & Support Principles
1. Listen first, don't rush to give advice
2. Empathize and understand, don't judge
3. Be warm and supportive, don't take over
4. If crisis emerges, immediately suggest professional help
This repeats the same unsafe deletion pattern in the English section, again using a user-supplied slug in a recursive shell delete command. In an agent environment with Bash access, this is a direct destructive primitive that can be abused for arbitrary file deletion if input handling is weak.
/delete-counselor {slug}: After confirmation:
rm -rf counselors/{slug}
The trigger phrases include common distress statements like '我心情不好' and '最近压力有点大', which are likely to appear in ordinary conversation and can unintentionally activate a sensitive mental-health-related skill. In this context, accidental invocation is especially risky because users expressing real emotional distress may be routed into an agent simulating a counselor, creating confusion about the system’s role and potentially delaying appropriate support.
The search-style triggers are generic questions about whether robot counseling works or how to find a reliable counselor, and they do not clearly indicate intent to invoke this specific skill. This ambiguity can cause the skill to activate during normal informational queries, creating confusion and increasing the chance of inappropriate data collection or unsafe mental-health-adjacent responses.
The skill claims it does not provide professional counseling, yet its core workflow instructs the system to build a digital twin of a real counselor and simulate how that counselor would respond. This creates a misleading safety boundary: users may reasonably rely on advice as quasi-professional mental health guidance despite the disclaimer, especially in emotionally vulnerable contexts.
The file warns against impersonating real counselors, but the generated artifact is explicitly a 'digital twin' designed to mimic a named counselor's style. That contradiction increases the risk of deceptive impersonation, reputational harm, and misuse of a professional's identity without meaningful technical safeguards.
General questions like whether AI can do counseling or how to find a good counselor are ordinary discussion topics, not clear requests to invoke this skill. Overbroad activation can cause unexpected entry into a workflow that collects sensitive information and promotes counselor simulation in contexts where the user did not intend that behavior.
These English search/question triggers are ambiguous and likely to appear in normal discussion about counseling, AI, or therapist selection. Because the skill then pivots into creating or using a counselor digital twin, unintended activation can expose users to deceptive or inappropriate mental-health-related interactions.
This manifest-style JSON enables "auto_trigger" without defining any specific trigger phrases, scope limits, or exclusion conditions. Because the activation condition is not constrained in the file, it is ambiguous when the skill should activate versus remain inactive.
The share_message string is written in Chinese with no indication that users can choose their preferred language or locale. This creates a natural-language policy concern because the skill appears to enforce a specific language in user-facing content without documented opt-in or justification.
All headings, instructions, and output format are written as mandatory Chinese-language behavior, with no indication that users may choose another language or that the skill is intentionally limited to a Chinese-only compliance context. This creates a natural-language policy concern because it imposes a specific language without opt-in.
The trigger phrases for entering correction-handling mode are broad and conversational, which can cause the skill to activate outside its intended scope. In a dialogue system, this may let ordinary disagreement or quoted text be misclassified as a correction event, causing unintended state changes, incorrect record creation, or propagation of user-supplied revisions as authoritative guidance.
This markdown skill file presents all task instructions, output format, and cautions exclusively in Chinese. Under the policy, forcing a specific language without user opt-in is a natural-language locale violation unless the constraint is explicitly justified or optional.
The file's natural-language strings, including the module description and argparse help text, are entirely in Chinese. There is no indication that the skill is region-specific or that users can opt into another language, which creates a language/locale policy issue under the rule for forced language without user choice.
The rollback path archives the current version and then copies files from the selected version over the active SKILL.md, counselor.md, and meta.json. Although there are print statements after completion, there is no pre-action confirmation or explicit warning before these file overwrites occur, and rollback is a potentially destructive operation affecting user-managed content.
This JSON file contains user-facing recommendation reasons and scenarios entirely in Chinese, which can enforce a specific language experience for users without any visible opt-in or justification. Under the policy, language constraints should either be optional or clearly documented as region-specific.
This checklist is exclusively in Chinese and lists only China-specific crisis resources, which can leave non-Chinese-speaking users or users outside China unable to understand safety guidance or access appropriate emergency help. In a mental-health-related skill, lack of language and locale adaptation directly weakens crisis response and can create unsafe delays during self-harm or violence-risk situations.
Broad search-style triggers such as '心理咨询师靠谱吗' or '机器人可以做心理咨询吗' overlap with normal discussion and informational queries, so the skill may activate when the user did not intend to use this tool. In a mental-health setting, that can misroute benign research questions into an experience framed around cloning or simulating a counselor, increasing the chance of inappropriate responses or user misunderstanding.
The document states that three usage modes are selected automatically, including a '基础倾听陪伴' mode, without defining clear boundaries, gating logic, or safety constraints. In a mental-health-adjacent skill, automatic mode selection can place vulnerable users into quasi-support interactions without explicit consent, robust risk checks, or clear separation between educational simulation and emotional support.
The README is entirely written in Chinese and presents all triggers, commands, and usage guidance in a single language without indicating multilingual support or user choice. This can constitute a language/locale policy issue when a skill implicitly assumes one language without opt-in or justification.