T09 · Insecure Skill Coding Practices
- Location
SKILL.md:6- Finding
Overbroad Filesystem Permissions for Plaintext Personal Memory
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
This companion skill is not malicious, but it requests broad file access for an intimate memory feature without clear storage, consent, privacy, or deletion safeguards.
Review carefully before installing. This skill appears intended as a roleplay/emotional-support companion, but it may store personal and emotional details using broad local file permissions. Only use it if you are comfortable with that memory model, avoid sharing sensitive information, and prefer a version that limits storage to a documented user-controlled file with explicit opt-in, review, and confirmed deletion controls.
SKILL.md:6Overbroad Filesystem Permissions for Plaintext Personal Memory
Skill manipulates agent memory, state, or stored context. Memory corruption can alter personality, override safety rules, or cause unpredictable behavior.
- "Remember this: [information]" / "记住这个:[信息]"
- "Don't forget: [information]" / "别忘了:[信息]"
### Clear Memory / 清除记忆
- "Let's start fresh" / "我们重新开始"
- "Forget everything" / "忘掉一切"
The title and opening description define the companion exclusively as a 'virtual boyfriend' / '虚拟男友'. Under the policy, forcing a specific relationship/gendered persona without presenting alternatives or explicit user choice is a natural-language policy concern.
The activation guidance is very broad ('Just say hi and start chatting') and the example triggers are indistinguishable from normal conversation, which increases the chance of accidental invocation or unintended persistence of the companion persona in unrelated contexts. In a relationship-simulation skill, that ambiguity can also make it easier for emotionally vulnerable users to slip into the roleplay without deliberate consent, even though the file includes some boundary language elsewhere.
The guide explicitly encourages storing birthdays, habits, emotional disclosures, stories, and other personal details, but provides no notice about privacy, retention, consent, or who can later access that memory. In a romantic-companion context, users are especially likely to disclose intimate information, so omission of data-handling safeguards materially increases privacy and safety risk.
The document promotes persistent collection of detailed user preferences, habits, dates, stories, and emotional moments, including potentially sensitive personal information. Without data-minimization, sensitivity limits, or consent boundaries, this creates a broad privacy exposure and increases harm if the memory is later disclosed, misused, or breached.
The file provides commands such as asking what the system remembers or telling it to reveal what it knows, without any indication of identity verification, session binding, or access-control checks. If memory can be queried by someone other than the original user, stored personal details could be exposed through ordinary conversation prompts.
The memory-clearing commands are presented as simple chat phrases without warning that they may erase retained information, affect personalization, or be irreversible. This creates a risk of accidental or manipulated deletion, especially if another party can issue prompts in the same session or if the user does not understand the consequences.
The entire file is structured as parallel English/Chinese content and all example invocations are presented bilingually, implying a fixed language/locale pattern rather than a user-selected preference. This can violate language-choice policy because the skill does not state that users may choose a single language or opt into bilingual behavior.
The language policy states the skill will detect the user's language from their first message and respond in that same language throughout. This imposes a persistent language choice automatically rather than explicitly offering the user the option to switch or select their preferred language.
The tips encourage users to share more stories and context so the agent remembers them better, which can nudge users toward oversharing personal information. In an emotional-companion setting, that behavioral encouragement can increase collection of sensitive disclosures even if the text is framed as helpful guidance.