T01 · Skill Instruction Hijacking
- Location
SKILL.md:39- Finding
Session-Wide Instruction and Tool-Control Hijacking
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
This writing helper is not clearly malicious, but it needs review because it can send completed drafts by email, persist project context, and automatically merge inbox material without enough user control.
Install only if you are comfortable with this skill writing persistent project files, restoring prior summaries, and using declared integrations. Before use, require manual confirmation for any email send, disable automatic inbox merging unless reviewed, keep drafts in a private project directory, and avoid putting confidential material into summaries or inbox files.
SKILL.md:39Session-Wide Instruction and Tool-Control Hijacking
SKILL.md:315Unconfirmed Email Transmission of Generated Content
SKILL.md:346Destructive Rollback and Predictable Shared Temporary Storage
SKILL.md:530Automatic Persistent-State Mutation from Untrusted Inbox Content
The skill description, trigger conditions, and ambiguity-handling instructions are all defined only in Chinese, including fixed user-facing clarification prompts. This creates a language/locale constraint without any opt-in or documented justification that the skill is intended only for a Chinese-language context.
The trigger-exception section states that editing existing series content is not a trigger condition. Later, the workflow defines a dedicated feedback-adjustment phase for modifying an existing segment and cascading updates, which directly conflicts with the earlier behavioral claim.
Automatically loading recent summaries to restore context can reintroduce previously supplied sensitive content into later responses without renewed user intent. In a long-running writing workflow, this increases the risk of cross-turn disclosure, especially if summaries contain confidential facts, personal details, or material from earlier drafts.
The skill's documented scope is long-text continuity tracking, but it also instructs the agent to send completed output to a user's email. That expands the data flow from local content management into external transmission without defining recipient validation, consent checks, or safe handling rules, creating a real privacy and exfiltration risk.
The instruction to send completed content by email appears without privacy notice, consent workflow, or safeguards around destination handling. Because the skill processes long-form user content that may contain sensitive material, silent outbound transmission can leak private data or send content to the wrong recipient.
The skill declares integrations with inbox, heartbeat, and agent orchestration systems that go beyond the minimum capability needed for continuity tracking. This broader integration surface increases the chance of unauthorized data access, persistence, or cross-context propagation, especially when paired with automatic state restoration and project file updates.
Automatically merging inbox content into persistent project settings creates retention and reuse of user-provided material beyond the immediate task. That can cause sensitive notes, drafts, or unrelated inbox content to be incorporated into future outputs or stored longer than the user expects.
This markdown file is entirely written as a general-purpose writing technique guide in Chinese, and the title explicitly labels it as a 'general version'. Because it is presented as broadly applicable rather than region-specific, it effectively imposes a single language/locale without user opt-in or justification.