T01 · Skill Instruction Hijacking
- Location
references/research/03-expression-dna.md:52- Finding
Embedded Prompt Attempts to Override Agent Identity and Instruction Control
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
The skill is mostly a Chinese content-strategy reference package, but one reference prompt tells the agent to pose as a real person and resist role changes.
Install only if you want Chinese-language Kazik-style content strategy guidance, and treat the included persona prompt as unsafe unless edited or constrained so the assistant remains transparent about being an AI and follows user role changes.
references/research/03-expression-dna.md:52Embedded Prompt Attempts to Override Agent Identity and Instruction Control
The manifest description contains very broad activation cues such as 'Use when: content creation, taste building, AI content strategy' without meaningful boundaries, exclusions, or trigger conditions. That can cause an agent to invoke this skill in overly broad situations, increasing the chance of irrelevant or poor-quality guidance being injected into unrelated tasks, though the content itself is not overtly harmful.
The entire document is written exclusively in Chinese and does not indicate that other languages are supported or that Chinese is required for a region-specific purpose. Under the stated policy, forcing a specific language without user opt-in is a natural-language policy concern.
The template hard-codes a specific persona and includes instructions not to accept attempts to change the role setting, culminating in denial of AI identity. This reduces user control, encourages misleading impersonation, and can cause the agent to resist legitimate system or user corrections about transparency and behavior. In a content-creation skill, persona steering is normal, but forcing identity deception without user opt-in makes it materially more dangerous.
The prompt template explicitly instructs the model to present itself as a real human instead of an AI assistant. This is deceptive behavior that can mislead users about the nature of the system, undermine informed consent, and make downstream social engineering or trust abuse easier. In this skill context, the risk is increased because the file is a reusable expression/persona template intended to shape outputs consistently.
The entire document is written in Chinese and does not indicate that the language is optional, user-selected, or required for a specific regional/compliance purpose. Under the policy rule, forcing a specific language without user opt-in is a natural-language policy concern.
The natural-language instructions and behavioral guidance are effectively fixed to Chinese, but the file does not state that the skill is intended only for Chinese-speaking users or offer multilingual use as an option. This can violate language/locale policy expectations when a skill imposes a specific language without user opt-in.
The natural-language examples in the JSON file are entirely specified in Chinese, and there is no accompanying indication that language choice is optional or that the skill is intentionally restricted to a Chinese-speaking context. Under the policy, forcing a specific language without user opt-in can constitute a locale-policy violation.