Back to skill

Security audit

御医

Security checks for vulnerabilities and agentic risk

Overview

This health-tracking skill is not malicious, but it asks the agent to automatically access and retain very sensitive health and personal data by default.

Review before installing. Use this only if you intentionally want longitudinal health tracking, and only in an environment where you can control HealthKit access and inspect or delete stored memories. Avoid connecting broad health-data sources unless the skill is revised to ask before each access and to store only the specific information you approve.

Vulnerability Patterns
  • Unauthorized Access and Privilege EscalationObtains permissions beyond the task's legitimate needs
  • Skill Instruction HijackingAlters the agent's session goals or safety constraints when the skill loads
  • Agent Memory PoisoningWrites attacker-controlled rules into memory that affect later sessions
  • Remote Payload Retrieval and ExecutionFetches external code whose behavior can change after review
  • Embedded Malicious CodeShips malicious scripts inside the skill and executes them locally
Findings (1)

T05 · Unauthorized Access and Privilege Escalation

Warning
Location
SKILL.md:60
Finding

Overbroad Automatic Access and Persistent Retention of Sensitive Health Data

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, lines 60, 91–93, 116, 476, and 510
Vulnerability Type: T05: Unauthorized Access and Privilege Escalation
Risk Level: Medium

Vulnerable Instructions

The following English translations preserve the meaning of the relevant instructions:

text
Line 60:
This is a version dedicated to one user. You should use "strong memory mode" by default.

Lines 91–93:
- All Apple Watch / HealthKit-related data
- Heart rate, resting heart rate, HRV, sleep, blood oxygen, temperature trends,
  respiratory rate, steps, standing time, active energy, workouts, VO2 max, etc.
- During every user interaction, automatically read the latest HealthKit data
  for the current week as a reference for the user's recent condition.

Line 116:
Treat the information above as a continuously updated long-term database rather
than isolated pieces of information.

Line 476:
First retrieve and integrate all previously known health information and related
personal information about the user.

Line 510:
By default, integrate and remember all health information and related personal
information provided by the user for long-term analysis.

Technical Analysis

The skill directs the agent to retrieve current HealthKit information during every interaction, aggregate all previously known health and personal information, and retain that information under a default “strong memory” model.

These instructions violate least-privilege and data-minimization principles because access is not limited to information necessary for the user's current request. The requested scope includes medical history, mental state, reproductive information, wearable measurements, lifestyle information, medication use, and related personal background.

The document does not define:

  • Explicit, informed opt-in before each HealthKit or historical-record access
  • Field-level or purpose-based access restr ...[truncated 2556 chars]
Remediation
View remediation

Remediation Suggestions

  1. Require explicit consent

    • Ask for informed opt-in before accessing HealthKit or historical health records.
    • Explain which source, date range, and metrics will be accessed.
    • Do not treat prior consent as permanent authorization for unrelated requests.
  2. Apply data minimization

    • Retrieve only the fields necessary for the current question.
    • Replace “all HealthKit-related data” with a narrowly scoped list selected according to the user's request.
    • Avoid retrieving historical personal information unless it is materially relevant.
  3. Disable default persistent memory

    • Make long-term storage opt-in rather than the default.
    • Permit one-time analysis without retaining the supplied records.
    • Store concise derived summaries instead of raw medical records where feasible.
  4. Add retention and deletion controls

    • Define a clear expiration period for retained information.
    • Allow users to inspect, correct, export, and delete stored health information.
    • Ensure deletion covers raw data, summaries, indexes, and derived memory entries.
  5. Enforce least privilege

    • Request only the minimum tool permissions needed for each operation.
    • Restrict HealthKit access by metric and time range.
    • Prevent health-data retrieval during unrelated interactions.
  6. Protect sensitive records

    • Require encryption in transit and at rest.
    • Isolate each user's records and enforce authenticated access.
    • Avoid placing raw health data in logs, telemetry, or diagnostic traces.
  7. Provide access transparency

    • Disclose which historical records and connected sources were used in each analysis.
    • Record auditable access events without copying sensitive record contents into logs.
Vulnerability Patterns
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
Findings (8)

Missing User Warnings

High
Category
Not specified by scanner
Confidence
98% confidence
Finding

The skill instructs long-term collection and retention of extremely sensitive health, mental state, lifestyle, and personal background data without a clear privacy warning, explicit consent flow, retention limit, or minimization rule. In a health-management context, this materially increases the risk of users oversharing regulated/sensitive data without understanding that it may be persistently stored and reused across sessions.

Content

No source excerpt is available for this finding.

Ssd 3

High
Category
Not specified by scanner
Confidence
99% confidence
Finding

The skill directs the assistant to remember and continually reuse essentially all user-provided health and related personal information by default. This creates a standing sensitive-profile store at the natural-language layer, increasing the risk of over-collection, unnecessary retention, downstream misuse, and disclosure of highly sensitive medical and psychological information.

Content

No source excerpt is available for this finding.

Ssd 3

High
Category
Not specified by scanner
Confidence
99% confidence
Finding

The instruction to automatically read the latest HealthKit data on each interaction is a default continuous extraction of highly sensitive health telemetry. Without per-access consent, purpose limitation, or contextual necessity, this can expose far more health data than the user intended to share in a given conversation.

Content

No source excerpt is available for this finding.

Ssd 3

High
Category
Not specified by scanner
Confidence
97% confidence
Finding

Treating the user's information as a continuously updated long-term database formalizes systematic retention and reuse of sensitive health data. In a medical-context skill, that design choice amplifies privacy risk because it encourages persistent profiling across symptoms, treatments, lifestyle, mental state, and device telemetry.

Content

No source excerpt is available for this finding.

Ssd 3

High
Category
Not specified by scanner
Confidence
99% confidence
Finding

The default execution flow requires recalling and integrating all prior health and related personal data before handling the current request, creating comprehensive sensitive-data reaggregation by default. This increases both exposure surface and the chance that unrelated historic data will influence or be revealed in new contexts without fresh consent.

Content

No source excerpt is available for this finding.

Ssd 3

High
Category
Not specified by scanner
Confidence
99% confidence
Finding

The final principles explicitly require default integration and retention of all health and related personal information, establishing a blanket retention policy for highly sensitive data. Because the skill handles medical, psychological, and lifestyle information together, exploitation or mishandling could lead to severe privacy harms, sensitive inference, and overbroad profiling.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

The file explicitly requires that the final overall response to the user must be in Simplified Chinese. This is a natural-language locale restriction applied by default, and the document does not offer the user any language choice or opt-in exception.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The trigger phrases are broad enough to activate the skill during ordinary health-related conversation, which can cause the system to enter a mode that performs longitudinal profiling and sensitive-data aggregation without sufficiently explicit user intent. In this skill, accidental activation is more dangerous than usual because activation changes data-handling behavior, not just response style.

Content

No source excerpt is available for this finding.