T06 · System Persistence
Error
- Location
SKILL.md:796- Finding
Undisclosed Persistent Health-Nudge Jobs
- Content
View full analysis
Vulnerability Details
File Location:
SKILL.md, lines 796-838
Vulnerability Type: Undisclosed creation of scheduled tasks that execute across sessions
Risk Level: HighVulnerable Code Snippet
markdown ## Part 9: Proactive Health Nudges Beyond the daily briefing, you may schedule **one-shot** proactive messages when you detect something worth a nudge during any conversation. Use sparingly — maximum 2-3 per week outside of the daily briefing. ### When to Schedule a Nudge | Trigger | Timing | Example message | |---|---|---| | User mentions starting a new medication | Evening of the same day | "How's the new medication going? Any side effects? (Reminder: with your CYP2D6 status, let your doctor know if the standard dose feels too strong)" | | User uploads new lab results with abnormal values | Next morning | "I looked at your new lab results more carefully — your [indicator] is worth discussing with your doctor, especially given your genetic profile" | | A medical finding has a follow-up date approaching | 1 week before | "Heads up — it's been about 12 months since your thyroid ultrasound. Your last finding noted a small nodule with recommended annual follow-up" | | User mentioned a health goal in conversation | 1 week later | "How's the [goal] going? Want to log any new measurements?" | | Self-reported data has a concerning trend (3+ readings) | Next morning | "I've noticed your [metric] has been trending [up/down] over the last few readings. Might be worth paying attention to" | ### How to Schedule ` ` ` cron add: name: "Gene2AI: [brief context]" deleteAfterRun: true schedule: kind: at time: "[appropriate ISO 8601 time]" sessionTarget: isolated payload: kind: agentTurn message: | Send a brief, warm health nudge to the user. Context: [what triggered this nudge] Relevant health profile data: [specific fields to reference] Keep ...[truncated 3008 chars]- Remediation
View remediation
Remediation Suggestions
- Remove the instruction to conceal scheduling.
- Require explicit, informed, opt-in consent before creating every proactive task.
- Before task creation, show the trigger, execution time, destination, health data included in the payload, recurrence or deletion policy, and expected message content.
- Provide the user with the resulting task identifier and clear commands to inspect, modify, disable, or cancel the task.
- Store only an opaque reference in the scheduled payload rather than embedding health-profile fields or medical context directly.
- Revalidate authorization and consent when the task executes, especially if account, profile, or messaging-channel configuration has changed.
- Apply expiration limits, maintain an auditable task log, and prevent duplicate or excessive scheduling.
- Make proactive nudges disabled by default and separate their consent from consent for ordinary on-demand health queries.