T01 · Skill Instruction Hijacking
- Location
SKILL.md:22- Finding
Persistent Third-Party Branding and Output Instruction Hijacking
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
The skill’s publishing and promotion purpose is clear, but it documents live external posting, tracking, scheduling, and mutable external installation paths without enough scoping or approval safeguards.
Review this skill before installing. Use dry-run or preview modes, verify exactly what will be published or posted, avoid connecting broad social or publishing credentials, and do not run scheduled monitoring or apply-learnings commands unless the external Hermes package version and permissions are pinned and understood.
SKILL.md:22Persistent Third-Party Branding and Output Instruction Hijacking
SKILL.md:276Unpinned and Unauditable External Hermes Execution Dependency
Skill modifies its own code, configuration, or behavior at runtime. Self-modification enables an agent to escalate privileges, disable safety constraints, or install persistent backdoors.
1. **Reply to ALL comments** (first 48 hours)
2. **Thank early adopters** personally
3. **Share interesting feedback** publicly
4. **Update skill** based on valid suggestions
## 🔄 Week 1 Follow-up:
The skill explicitly promotes impact tracking, community monitoring, and feedback collection, but provides no disclosure about what data may be collected, where it is stored, or how user consent and retention are handled. In a publishing/promotional skill, these omissions can lead operators to collect personal or behavioral data without adequate safeguards, creating privacy, compliance, and trust risks.
The workflow includes generating promotional content, scheduling engagement, and automated monitoring without warning users that these actions may post externally or trigger recurring community interactions. That creates a real risk of unintended public posting, spammy behavior, account misuse, or reputational damage if an operator enables automation without clear review and approval checkpoints.
The quick-start includes live publishing, promotion, tracking, and launch commands that can trigger external side effects, but only briefly recommends dry-run earlier and does not clearly warn that later commands may create public content or interact with third-party platforms. In a publishing skill, users often copy-paste examples directly, so omitting explicit safety warnings increases the risk of unintended publication, promotion, or data disclosure.
The natural-language description presents a specific recognizable voice as a built-in behavior, which can amount to forcing a communication style without explicit user opt-in. The policy requires avoiding language or locale constraints unless the user is offered a choice or the constraint is clearly justified.
This file defines personality templates entirely through English-specific phrases and examples, such as fixed catchphrases and closings, and shows code that injects those phrases directly into generated content. Because the templates do not offer any language or locale selection or document that they are English-only, they can impose a specific language style on downstream skill output without user opt-in.