T01 · Skill Instruction Hijacking
- Location
SKILL.md:10- Finding
Session-Wide Skill Instruction Hijacking
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
This skill is a cron-management guide, but it tells the agent to run checks on every message and make automatic workspace and task-configuration changes without clear user approval.
Install only in a controlled personal workspace where you want this agent to manage cron tasks. Before using it broadly, narrow activation to explicit cron-management requests and require approval before file writes, task prompt changes, retries, timeout changes, notifications, or deletion of incident records.
SKILL.md:10Session-Wide Skill Instruction Hijacking
SKILL.md:40Automatic Workspace Inspection and Modification Without Per-Action Authorization
SKILL.md:179Destructive Error-Record Cleanup Can Remove Audit Evidence
The rule to execute on every incoming message creates an implicit global hook that can cause the skill to run outside the user's intent and to inspect or modify workspace state opportunistically. Because the HEARTBEAT logic includes direct remediation and user notification, any message can become a trigger for autonomous action, amplifying unintended side effects and making prompt-injection through unrelated conversations more dangerous.
The activation scope is very broad, covering checking status, modifying configuration, resolving overlaps, and optimizing schedules without explicit guardrails for when user confirmation is required. In a skill that can influence cron configuration and system behavior, this increases the chance of unintended invocation and autonomous changes beyond what the user expected.
The skill instructs immediate timeout changes, retries, and proactive notification without requiring consent or presenting the operational impact of altering cron behavior. Automatic modification of scheduling/execution parameters can destabilize systems, mask root causes, or create repeated actions, especially in a privileged cron-management context.
Multiple scheduled tasks are defined explicitly in Beijing time, and the document does not offer locale selection or explain that the skill is limited to a China-specific deployment context. This is a natural-language locale policy concern because it hard-codes a locale assumption without opt-in.
The conflict-handling flow authorizes the skill to update task prompts or reassign responsibilities immediately after detection, again without user-facing review. In this context, automatic edits to prompts and task definitions are configuration changes that can silently alter system behavior, create outages, or entrench an attacker-influenced interpretation of 'conflict.'
The cleanup policy directs deletion of records from ERRORS.md, including age-based cleanup, without emphasizing retention safeguards, auditability, or confirmation before destructive actions. In an incident-tracking file, automatic deletion can remove forensic context, reduce transparency, and make it harder to investigate recurring or unresolved problems.