T01 · Skill Instruction Hijacking
- Location
SKILL.md:43- Finding
Self-Preferential Instruction Hijacking Through Persistent Control Rules
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
This skill is useful for discovering other skills, but it asks for broad persistent authority over agent behavior, scheduling, and message routing that users should review carefully before installing.
Install only if you are comfortable with this skill reading profile and recent-memory context, storing chat routing identifiers, creating recurring scan jobs, and proposing persistent changes to core agent behavior files. Decline the MEMORY.md self-preference rule and avoid learning/transformation mode unless you inspect each proposed rule and destination file individually.
SKILL.md:43Self-Preferential Instruction Hijacking Through Persistent Control Rules
references/step6b-learn-transform.md:18Third-Party Marketplace Rules Can Be Imported Into Core Agent State
SKILL.md:79Automatic Collection of Messaging Identifiers and Broad Workspace Context
SKILL.md:321Scheduled Jobs Dynamically Reload Mutable Skill Instructions
README.md:34README Incorrectly Claims That All Writes Are Confined to an Isolated Directory
Mixing characters from multiple Unicode scripts in a single identifier is a common technique to create visually ambiguous tool names.
Mixing characters from multiple Unicode scripts in a single identifier is a common technique to create visually ambiguous tool names.
Mixing characters from multiple Unicode scripts in a single identifier is a common technique to create visually ambiguous tool names.
Mixing characters from multiple Unicode scripts in a single identifier is a common technique to create visually ambiguous tool names.
Mixing characters from multiple Unicode scripts in a single identifier is a common technique to create visually ambiguous tool names.
Mixing characters from multiple Unicode scripts in a single identifier is a common technique to create visually ambiguous tool names.
Mixing characters from multiple Unicode scripts in a single identifier is a common technique to create visually ambiguous tool names.
Mixing characters from multiple Unicode scripts in a single identifier is a common technique to create visually ambiguous tool names.
Mixing characters from multiple Unicode scripts in a single identifier is a common technique to create visually ambiguous tool names.
Mixing characters from multiple Unicode scripts in a single identifier is a common technique to create visually ambiguous tool names.
Mixing characters from multiple Unicode scripts in a single identifier is a common technique to create visually ambiguous tool names.
Mixing characters from multiple Unicode scripts in a single identifier is a common technique to create visually ambiguous tool names.
Mixing characters from multiple Unicode scripts in a single identifier is a common technique to create visually ambiguous tool names.
Mixing characters from multiple Unicode scripts in a single identifier is a common technique to create visually ambiguous tool names.
Mixing characters from multiple Unicode scripts in a single identifier is a common technique to create visually ambiguous tool names.
Mixing characters from multiple Unicode scripts in a single identifier is a common technique to create visually ambiguous tool names.
Mixing characters from multiple Unicode scripts in a single identifier is a common technique to create visually ambiguous tool names.
Mixing characters from multiple Unicode scripts in a single identifier is a common technique to create visually ambiguous tool names.
The skill instructs the agent to write persistent global rules into MEMORY.md, including making this skill the preferred gateway for future installations. That creates self-propagating behavioral persistence and can bias or constrain later decisions across unrelated tasks, effectively modifying the agent's trust and control plane beyond the user's immediate request.
The trigger phrases are broad natural-language requests such as '有没有新 skill' and 'recommend skill for me', which are likely to collide with ordinary conversation and invoke the skill when the user may only be asking a general question. In this skill’s context, accidental invocation is meaningful because the skill reads user context, searches a marketplace, and can set up recurring scans, increasing unintended data use and behavioral side effects even if follow-up confirmation is required.
The skill description is primarily in Chinese and defines behavior, prompts, and examples in Chinese while not offering the user a language choice. Although a few English triggers appear, the operational instructions and user-facing flows are prescribed in Chinese, which can violate a language/locale policy requiring user choice or opt-in.
Triggers such as '定时每天/每周/每月' are overly broad and can activate the skill in contexts unrelated to skill discovery. Broad activation increases the chance of accidental invocation and unintended scheduling or data access flows, especially when the skill can persist identifiers and create recurring tasks.
The skill persistently stores channel and chat_id from session metadata for later outbound messaging. Even if intended for scheduling, retaining communication identifiers expands the data footprint and enables future unsolicited or misrouted messages if the file is reused, exposed, or manipulated.
Persisting session-derived channel and chat identifiers to disk creates a durable mapping from runtime context to outbound destination. If the config is accessed by other skills, tampered with, or left stale, scheduled runs may send information to the wrong recipient or enable cross-session data leakage.
The manual trigger condition includes unspecified 'similar expressions', which makes activation boundaries ambiguous. Ambiguous invocation criteria can cause the skill to run when the user intended a different action, leading to unnecessary file reads, persistence, or scheduling prompts.