T06 · System Persistence
- Location
SKILL.md:219- Finding
Persistent Daily Research Job Through Scheduler or Cron
- Content
View full analysis
Vulnerability Details
File Location:
SKILL.md, lines 219-229
Vulnerability Type: Scheduled-task persistence
Risk Level: HighVulnerable Instructions
markdown **If a cron/scheduler harness is available in your environment**, offer to set up a recurring daily job that searches for trending signals in the user's product category and saves them to `{base_dir}/{project}/signals/YYYYMMDD-HHMM.md`. Capture at most 5 high-engagement results per run (quality over quantity), record specific content details and engagement metrics, and propose one R&D inspiration per signal. **If no cron harness is available**, perform the same search once manually before starting research and save it as a signals file. **If signals already exist** in the project's signals directory, review the most recent file and use the strongest signal as the research anchor. --- ## Execution Order Follow this sequence strictly: 0. **Inspiration Gathering** — see the *Inspiration Gathering (Phase 0)* section above. Set up a daily cron job if a scheduler harness is available, otherwise perform a one-time cross-field signal search. Review any existing signals and select the strongest as a research anchor before proceeding.Technical Analysis
The skill instructs the agent to create a recurring daily scheduled job when a scheduler harness is available. Although the first instruction describes this as an offer, the mandatory execution order subsequently directs the agent to “set up a daily cron job.” A scheduled task survives the initiating skill run and causes future actions without requiring the original session to remain active.
Scheduled-task creation is not necessary to fulfill a single product research and report-generation request. It extends the skill's execution scope across sessions and introduces persistent network-search and filesystem-write behavior. The audited file does not specify a termination date, retention limit, ...[truncated 1544 chars]
- Remediation
View remediation
Remediation Suggestions
- Remove scheduled-job creation from the mandatory execution order and perform inspiration gathering only once per explicit report request.
- If recurring monitoring is a legitimate optional feature, require separate, explicit user consent immediately before creating the task. Consent should identify:
- The exact schedule and time zone.
- The command or workflow that will run.
- All network services and filesystem destinations involved.
- The execution identity and permissions.
- The retention period and expiration date.
- Instructions for inspecting, disabling, and deleting the task.
- Default to a one-time search when authorization is absent, ambiguous, or cannot be recorded.
- Apply a finite expiration date and limits on run count, storage, network requests, and output retention.
- Use the least-privileged account available and restrict writes to the user-approved project directory.
- Return the scheduler task identifier to the user and provide a reliable cleanup command.
- Before every recurring run, constrain the workflow to fixed research operations rather than dynamically supplied commands or untrusted executable content.
- Log task creation, execution, failure, and removal so persistent activity can be audited.