Back to skill

Security audit

Fashion Designer

Security checks for vulnerabilities and agentic risk

Overview

The skill is a coherent fashion assistant, but it stores personal style/profile data and syncs it to Feishu without clear consent or retention controls, and it gives unsafe image-download instructions.

Review before installing. This skill can build a lasting local profile from gender, age range, body type, style preferences, and feedback, and may put that information into Feishu documents. Use it only if you are comfortable with that storage and sharing, and prefer a version that asks before saving profiles, supports deletion, limits Feishu document contents, and validates downloaded image URLs safely.

Vulnerability Patterns
  • Insecure Skill Coding PracticesFinds exploitable flaws such as hardcoded secrets or command injection
  • 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 (2)

other

Warning
Location
SKILL.md:22
Finding

Persistent Collection and External Disclosure of Personal Profile Data

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, lines 22-27, with related behavior at lines 31-36 and 58-65
Vulnerability Type: Privacy-invasive data collection and disclosure
Risk Level: Medium

Vulnerable Code

markdown
1. **User profile first**:Before each recommendation, gather user info (gender, age, body type, style preference) and save to USER.md; carry profile into all future recommendations
2. **Real-time progress updates**:At the start of each step, send a brief IM message updating the user (e.g., "Doing X, please wait...") with a playful, flexible tone
3. **Message + document sync**:Deliver recommendations via IM card and IM document simultaneously
4. **No duplicate document per session**:Create only one recommendation document per session(keyed by session ID), then update it
5. **Strict image filtering**:Max 3 images per outfit; try each candidate image; skip if upload fails; if all fail, remove image—no placeholders
6. **User feedback loop**:Based on feedback, update user profile and style preferences in USER.md; summarize if content exceeds 200 characters. If IM is Feishu, update the Feishu document—highlight liked outfit titles, remove disliked ones, keep unmentioned ones.

Related profile collection and external-document instructions include:

markdown
If user info is available from chat history or memory, use it; otherwise politely ask:
- Gender
- Age range (child / youth / middle-aged / senior)
- Body type
- Style preference
markdown
### First Creation
- Create Feishu document with user profile and all recommended outfits

Technical Analysis

The Skill requires the Agent to collect gender, age range, body type, and style preferences, persist these attributes in USER.md, and reuse them in future recommendations. It further directs the Agent to copy the profile into an external Feishu document.

Persistent storage and external synchronization are ...[truncated 1751 chars]

Remediation
View remediation

Remediation Suggestions

  1. Make persistent profile storage explicitly opt-in and explain which attributes will be retained.
  2. Default to session-only processing when the user has not consented to persistence.
  3. Request separate confirmation before copying profile attributes into Feishu.
  4. Minimize collected data; do not require age or body type when the current request can be answered without them.
  5. Avoid extracting profile attributes from unrelated chat history or memory unless the user explicitly asks for reuse.
  6. Exclude personal profile details from external documents by default. Include only recommendation content unless the user approves each category of profile data.
  7. Define retention limits and provide commands for users to inspect, correct, and delete stored profile information.
  8. Validate Feishu document permissions before upload and use the narrowest possible sharing scope.
  9. Record consent state separately so that a prior styling request is not treated as indefinite authorization for storage or external transmission.

T09 · Insecure Skill Coding Practices

Error
Location
SKILL.md:85
Finding

Unvalidated Remote Image Download Enables SSRF and Unsafe Temporary-File Handling

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, lines 85-107
Vulnerability Type: Unvalidated network retrieval and unsafe temporary file usage
Risk Level: High

Vulnerable Code

markdown
### Correct Process
1. **Download image**:Use curl to download product image to `/tmp/` 
   ```bash
   curl -s -o /tmp/img1.jpg "iamge URL"
   ```

2. **Upload to Feishu**:Use feishu_doc_media tool to upload local image
   ```json
   {
     "action": "insert",
     "doc_id": "document ID",
     "file_path": "/tmp/img1.jpg",
     "type": "image"
   }
   ```

3. **Get token**:Upload success returns file_token (e.g., V6vfbBRImojQ9KxQCKpcq1FLnAb)

4. **Insert into document**:Use feishu_update_doc with insert_after mode to insert image under corresponding outfit title

5. **Delete temporary file**:Remove downloaded image from `/tmp/` after processing to avoid excessive disk usage

Technical Analysis

The image URL originates from browsed product content and is supplied to curl without validation. The instructions do not restrict the URL scheme or destination host, block loopback and private network ranges, prevent access to cloud metadata services, constrain redirects, enforce download-size or timeout limits, or verify that the response is a valid image.

A malicious or compromised shopping page could present an image URL targeting an internal service, local network endpoint, or cloud metadata address. The Agent would then make the request from its own network context. The response could subsequently be uploaded to Feishu, creating a possible data-exfiltration path.

The fixed path /tmp/img1.jpg introduces additional weaknesses. Another local process could pre-create that path as a symbolic link, potentially causing the command to overwrite a file writable by the Agent. Concurrent Skill runs can also overwrite or upload each other's content. Deletion after processing does not prevent races and is ...[truncated 1868 chars]

Remediation
View remediation

Remediation Suggestions

  1. Permit only https image URLs and reject unsupported schemes such as file, ftp, and other protocol handlers.
  2. Allowlist trusted image hosts associated with the configured shopping sites rather than accepting arbitrary URLs extracted from page content.
  3. Resolve destination hostnames before each request and reject loopback, link-local, private, multicast, reserved, and cloud-metadata address ranges for both IPv4 and IPv6.
  4. Disable redirects or validate every redirect target against the same scheme, hostname, and resolved-address policy.
  5. Apply strict connection, transfer, and total-operation timeouts.
  6. Enforce a conservative maximum response size before and during download.
  7. Require an expected image MIME type and decode the file with a trusted image library before upload. Do not rely on the filename extension or server-provided content type alone.
  8. Create a unique temporary directory and file using secure random creation with restrictive permissions. Do not use a predictable shared path such as /tmp/img1.jpg.
  9. Open temporary files with protections against symbolic-link following and accidental overwrite.
  10. Keep the file descriptor or securely bind the validated file to the upload operation to reduce time-of-check/time-of-use races.
  11. Guarantee cleanup with a finally-style mechanism even if download, validation, or upload fails.
  12. Do not upload downloaded content to Feishu until host validation, size validation, MIME verification, and image decoding have all succeeded.
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
Findings (6)

Missing User Warnings

High
Category
Not specified by scanner
Confidence
97% confidence
Finding

The skill instructs collecting sensitive preference/profile attributes such as gender, age range, body type, and style preference, then persisting them to USER.md without consent, retention disclosure, or minimization. This creates a durable profile that could be reused across sessions in ways the user does not expect, increasing privacy and compliance risk.

Content

No source excerpt is available for this finding.

Missing User Warnings

High
Category
Not specified by scanner
Confidence
96% confidence
Finding

The skill continuously updates persisted preference data based on feedback, effectively performing ongoing behavioral profiling without a clear disclosure. Because this modifies long-term memory from conversational reactions, users may unknowingly train a durable profile that influences future outputs and could expose personal preferences if accessed later.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
89% confidence
Finding

The top-level description is broad enough that the skill may activate for general fashion conversation, not just explicit requests for outfit recommendations. Over-broad triggering can cause unintended collection of profile data and initiation of external browsing or document-writing behaviors without clear user intent.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The 'When to Use' criteria are ambiguous and effectively cover any scenario involving clothing or style discussion. In this skill, that matters because activation leads to profiling, persistence to USER.md, external scraping, and possible Feishu document creation, making accidental invocation more privacy-sensitive than a purely local advisory skill.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
88% confidence
Finding

The workflow directs the agent to fetch product and image data from external sites and potentially use external browser/scraping tooling, but does not require notifying the user that network requests will occur. This can surprise users, leak context to third-party sites, and expand the attack surface through external content retrieval.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The skill describes creating and syncing recommendation content into Feishu documents without clearly warning users that their profile and recommendations may be stored in an external collaboration service. Since the document can contain personal profile details and shopping preferences, silent synchronization creates avoidable privacy and data-governance risk.

Content

No source excerpt is available for this finding.