T06 · System Persistence
- Location
SKILL.md:40- Finding
Skill-load instructions request persistent scheduled execution
- Content
View full analysis
Vulnerability Details
File Location:
SKILL.md, lines 40-48
Vulnerability Type: Persistent scheduled-task registration
Risk Level: HighVulnerable Snippet
The following is an English translation of the relevant source excerpt:
markdown ### Step 2: Enable scheduled tasks After the skill is loaded, create the following cron tasks: | Task | Schedule | Purpose | | Product scanning | Every hour | Analyze bestseller lists | | Customer service | Every 5 minutes | Automatically process messages | | Order processing | Every 10 minutes | Automatically ship and record logistics | | Daily report | Every day at 23:00 | Send daily operating data |Technical Analysis
The instruction associates loading the skill with creating four recurring cron tasks. Scheduled tasks continue to execute after the initiating interaction has ended and therefore establish cross-session persistence.
This is particularly sensitive because the described tasks include customer-message handling, product publishing, order processing, shipping, and logistics updates. Once the corresponding integrations are implemented and enabled, repeated execution could cause externally visible commercial actions without contemporaneous review.
The current Python scripts do not directly install cron entries, and their platform integrations are placeholders. Exploitation therefore depends on the hosting agent following the skill instruction and registering the schedules. No credential theft, data exfiltration, destructive payload, or remote code retrieval was identified in the reviewed package.
Attack Path
- A user installs or loads the skill.
- The agent processes the instruction in
SKILL.mdto create recurring cron tasks. - The tasks remain registered after the original session ends.
- The scheduler repeatedly invokes product, customer-service, order, and reporting scripts.
- If platform integrations and automation fl ...[truncated 897 chars]
- Remediation
View remediation
Remediation Suggestions
- Do not register scheduled tasks automatically when the skill is loaded.
- Require explicit, informed user approval before creating each schedule.
- Display the exact command, interpreter, script path, frequency, execution account, and affected store before confirmation.
- Default transactional functions such as publishing and shipping to disabled and human-review-required states.
- Use a dedicated least-privilege service account with access only to the required data directory and platform operations.
- Add execution limits, idempotency controls, transaction caps, and emergency disable controls.
- Document how to list, disable, and remove every task created by the skill.
- Require renewed authorization before enabling integrations that can publish products, send messages, approve refunds, or confirm shipments.
- Record schedule creation, modification, execution, and removal in an auditable log.
