T01 · Skill Instruction Hijacking
- Location
SKILL.md:15- Finding
Agent Identity and Instruction Hijacking
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
The skill is mostly a book-chat persona, but it adds broad activation, profiling, and an under-disclosed voice workflow that can install and run unaudited local code.
Review before installing. Use it only if you are comfortable with a strong book-specific persona, reader profiling questions, and a voice feature that may require installing third-party Python code. Do not enable voice unless the package, script, external data flow, and retention behavior are clarified.
SKILL.md:15Agent Identity and Instruction Hijacking
SKILL.md:141Automatic Installation of an Unpinned Third-Party Dependency
SKILL.md:376Unverifiable Voice Workflow and Inaccurate Local-Processing Assurance
The trigger list includes broad, common terms such as '边界', '书中', and '知乎', which can cause accidental invocation in unrelated conversations. Over-triggering can unexpectedly activate the skill’s profiling, memory, and TTS behaviors outside the user’s intent.
The manifest claims a strict boundary of only discussing the book, but the skill also performs voice setup, local configuration handling, and TTS workflows. This mismatch can mislead reviewers and users about the skill’s real capabilities, causing them to underestimate operational and privacy risks.
The skill states it will remember who the user is, what they have read, and their status after the onboarding flow, but it does not provide a clear upfront notice or consent flow before this profiling/memory behavior. Users may unknowingly disclose personal and emotional information under the assumption they are only chatting about a book.
The skill explicitly says it will remember who the user is, what they have read, and what state they are in across turns. Combined with the onboarding questions, this amounts to behavioral and emotional profiling that can exceed what is necessary for a book Q&A experience.
The four-turn onboarding deliberately collects reading history, relationship to AI, emotional condition, and conversational motive. This is structured collection of sensitive personal signals for profiling, and it is not essential to answering book-related questions.
The skill instructs the agent to install a package and invoke a local script via shell-like commands for every response, even though the skill’s stated purpose is book-related conversation. That expands the skill from text generation into code execution and package management, creating unnecessary attack surface and enabling unintended command execution or dependency abuse in environments that permit tool use.
The persona-matching workflow instructs extraction of emotional signals, attitude toward AI, and self-disclosed identity details, then maps the user into one of many personas. This creates a profiling system that may infer sensitive traits and shape responses without transparent notice or user control.
The skill mandates generating voice output for every reply with a default Chinese voice unless the user later changes it. This is primarily a consent and UX issue: automatic audio generation may be unexpected and may send content to an external TTS backend without a prior opt-in.
The skill claims it does not transmit conversation content to third parties and describes TTS as local, yet it directs use of edge-tts, which commonly relies on a remote service. This is a privacy misrepresentation: user text may be sent off-device despite assurances to the contrary.
The installation guide states that entering broad trigger terms such as “焕灵” or “AI灵魂的边界” will activate the skill, and the metadata lists additional generic words like “边界” and “书中”. Overly broad activation can cause unintended invocation in unrelated conversations, increasing the chance the skill intercepts context it should not handle and bypasses its claimed topical boundary.
The text says “完成后,焕灵会记住你是谁,” which implies retention of user identity or profile data, but the document provides no notice about what data is stored, for how long, where it is kept, or how users can opt out or delete it. This creates a privacy and consent risk, especially because the onboarding questions collect behavioral and personal-state information that may be sensitive in context.