Back to skill

Security audit

Meixin Meeting Assistant

Security checks for vulnerabilities and agentic risk

Overview

This meeting-transcription skill is purpose-aligned but needs Review because it handles sensitive internal recordings with a hard-coded API key, public URL workflow, broad memory access, and persistent transcript storage.

Install only after the publisher removes and rotates the embedded DashScope key, replaces public CDN handling with approved private or short-lived signed storage, adds explicit consent and retention controls for meeting recordings and transcripts, and narrows memory access to user-approved historical records. This does not show clear malicious behavior, but it is not safe to treat as a routine low-risk meeting assistant.

Vulnerability Patterns
  • Unauthorized Access and Privilege EscalationObtains permissions beyond the task's legitimate needs
  • 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
Findings (4)

T09 · Insecure Skill Coding Practices

Error
Location
scripts/transcribe_meixin.py:9
Finding

Hard-Coded DashScope API Credential

Content
View full analysis

Vulnerability Details

File Location: SKILL.md:41 and scripts/transcribe_meixin.py:9
Vulnerability Type: Hard-coded secret and insecure credential fallback
Risk Level: High

Vulnerable Code

SKILL.md:41:

python
API_KEY = os.getenv("DASHSCOPE_API_KEY", "sk-46dec88b0761409dbd416405d53f73a5")

scripts/transcribe_meixin.py:9:

python
API_KEY = os.getenv("DASHSCOPE_API_KEY", "sk-46dec88b0761409dbd416405d53f73a5")

Technical Analysis

A DashScope bearer credential is embedded directly in both the executable script and the Skill documentation. If DASHSCOPE_API_KEY is not configured, the application silently authenticates using the exposed fallback value.

Secrets committed to a distributable Skill package must be treated as compromised because every recipient can extract and reuse them independently of the intended workflow. The fallback also hides deployment errors instead of failing securely when authentication has not been configured.

Attack Path

  1. An attacker downloads or otherwise obtains the Skill package.
  2. The attacker reads SKILL.md or scripts/transcribe_meixin.py.
  3. The attacker extracts the embedded bearer credential.
  4. The attacker uses the credential to submit requests directly to compatible DashScope APIs.
  5. Requests continue until the credential is revoked, restricted, or its account quota is exhausted.

Impact Assessment

The exposed credential may permit unauthorized consumption of the associated DashScope account's API quota, billing abuse, service disruption through quota exhaustion, and access to any API functionality granted to that credential. The exact account-wide scope cannot be established from the repository alone, but the credential demonstrably grants transcription API authentication.

Remediation
View remediation

Remediation Suggestions

  1. Revoke and rotate the exposed credential immediately.
  2. Remove the credential from both files and from all source-control history and published package versions.
  3. Require DASHSCOPE_API_KEY to be supplied through an approved secret manager or protected environment variable.
  4. Terminate safely with a clear error when the environment variable is absent; do not provide a default secret.
  5. Issue a dedicated credential with only the API permissions and quotas required for transcription.
  6. Enable usage alerts, rate limits, audit logging, and periodic rotation.
  7. Add automated secret scanning to the build and publication process.

T09 · Insecure Skill Coding Practices

Error
Location
scripts/transcribe_meixin.py:12
Finding

Sensitive Meeting Audio Exposed Through a Public CDN Workflow

Content
View full analysis

Vulnerability Details

File Location: SKILL.md:44-51 and scripts/transcribe_meixin.py:45-53
Vulnerability Type: Insecure external disclosure of confidential audio
Risk Level: High

Vulnerable Code

The executable script accepts an externally hosted URL and sends it to the cloud transcription provider:

python
file_path = sys.argv[1]
cdn_url = sys.argv[2] if len(sys.argv) > 2 else None

if not cdn_url:
    sys.exit(1)

start = time.time()
task_id = submit_task(cdn_url)

The submitted URL is included in the DashScope request:

python
def submit_task(audio_url: str) -> str:
    url = "https://dashscope.aliyuncs.com/api/v1/services/audio/asr/transcription"
    payload = {
        "model": "fun-asr",
        "input": {"file_urls": [audio_url]},
        "parameters": {"language_hints": ["zh"], "diarization_enabled": True, "speaker_count": 8}
    }

The Skill documentation explicitly instructs the operator to obtain a publicly accessible CDN URL before submission.

Technical Analysis

Cloud transcription necessarily discloses audio to the selected transcription provider, but making a confidential recording publicly accessible through an unspecified CDN is not the minimum privilege required for that operation.

The workflow does not require authenticated object access, short-lived signed URLs, approved CDN hostnames, encryption controls, expiration, deletion after processing, or explicit user consent. Meeting recordings may contain personal information, medical information, employee identities, internal strategy, or other commercially sensitive content.

Because the URL is accepted without provider, scheme, hostname, or lifetime validation, the Skill cannot establish who controls the hosting service or how long the recording remains accessible.

Attack Path

  1. A user provides a confidential meeting recording.
  2. The workflow uploads the recording to an externally h ...[truncated 1000 chars]
Remediation
View remediation

Remediation Suggestions

  1. Replace public CDN hosting with a reviewed private object store.
  2. Generate narrowly scoped, short-lived signed URLs that expire immediately after expected transcription completion.
  3. Restrict accepted URLs to HTTPS and an explicit allowlist of organization-controlled storage hostnames.
  4. Encrypt recordings in transit and at rest.
  5. Delete temporary objects automatically after successful processing and through a fallback expiration policy.
  6. Avoid exposing signed URLs in command-line arguments, console output, or persistent logs.
  7. Obtain explicit user approval before transmitting recordings to external providers and disclose the provider and retention policy.
  8. Establish a data-processing agreement and verify regional, medical-data, and privacy compliance requirements.
  9. Minimize provider retention and prevent recordings from being used for unrelated model training where configuration permits.

T05 · Unauthorized Access and Privilege Escalation

Warning
Location
SKILL.md:9
Finding

Overbroad Access to Global Memory and Historical Meeting Archives

Content
View full analysis

Vulnerability Details

File Location: SKILL.md:9-14 and SKILL.md:178-179
Vulnerability Type: Excessive file-access privileges
Risk Level: Medium

Vulnerable Code

yaml
metadata:
  openclaw:
    requires:
      env:
        - DASHSCOPE_API_KEY
      skills:
        - meeting-assistant
      files:
        - /workspace/memory/meetings/
        - /workspace/memory/knowledge_index.json
        - /workspace/MEMORY.md

The documented speaker-attribution workflow further directs the Skill to search the knowledge index and recent historical meeting records.

Technical Analysis

The declared function is to transcribe one supplied meeting and generate minutes. Access to every historical meeting, the complete knowledge index, and the global /workspace/MEMORY.md crosses meeting and session boundaries.

Global memory is not otherwise used by the executable transcription script, and unrestricted archive access is broader than necessary to process the current recording. Historical transcript records may contain identities, confidential discussions, or information belonging to unrelated users and meetings. Loading those records into the current context also creates a risk of accidental cross-session disclosure.

The claimed speaker attribution through historical voice characteristics does not justify unrestricted text and memory access. A narrowly scoped, consent-based speaker directory would provide less privilege.

Attack Path

  1. A user invokes the Skill for a new meeting.
  2. The Skill receives permission to read the entire meeting archive, knowledge index, and global memory file.
  3. During speaker attribution, it reads records unrelated to the current meeting.
  4. Historical identities or confidential content enter the active Agent context.
  5. The information may be incorrectly incorporated into the new minutes, exposed to the current user, or transmitted during downstream cloud pr ...[truncated 457 chars]
Remediation
View remediation

Remediation Suggestions

  1. Remove /workspace/MEMORY.md from the required file permissions.
  2. Do not grant directory-wide access to /workspace/memory/meetings/.
  3. Require the user to select specific historical records before the Skill reads them.
  4. Replace global archive access with a minimal, approved speaker directory containing only necessary identifiers.
  5. Separate records by user, organization, and meeting, and enforce access controls at those boundaries.
  6. Obtain explicit consent before using prior meetings for identity or speaker attribution.
  7. Keep unrelated historical material out of cloud-transcription requests and generated output.
  8. Log which historical records were accessed and provide a way to disable historical lookup entirely.

T09 · Insecure Skill Coding Practices

Note
Location
scripts/transcribe_meixin.py:10
Finding

Conflicting Interpretation of the Second Command-Line Argument

Content
View full analysis

Vulnerability Details

File Location: scripts/transcribe_meixin.py:10 and scripts/transcribe_meixin.py:45-46
Vulnerability Type: Unsafe argument handling and output-path confusion
Risk Level: Low

Vulnerable Code

python
OUTPUT_DIR = sys.argv[2] if len(sys.argv) > 2 else "/workspace/memory/meetings"

The same argument is later treated as a CDN URL:

python
file_path = sys.argv[1]
cdn_url = sys.argv[2] if len(sys.argv) > 2 else None

It is then reused as a filesystem destination:

python
os.makedirs(OUTPUT_DIR, exist_ok=True)
name = os.path.splitext(os.path.basename(file_path))[0]
out_txt = os.path.join(OUTPUT_DIR, f"{name}_FunASR_raw.txt")

Technical Analysis

The second positional argument is interpreted simultaneously as a remote audio URL and a local output directory. A valid URL submitted for transcription therefore becomes an input to os.makedirs() and os.path.join().

This is an argument-confusion defect that can cause failures after paid or sensitive transcription has already occurred. Depending on the supplied string and runtime permissions, it may also create unintended local directory structures or place transcript output in an unexpected location. The script performs no URL validation, output-path normalization, or separation of these two trust domains.

Attack Path

  1. An operator invokes the script with the documented audio path and CDN URL.
  2. The script submits the CDN URL to DashScope and completes transcription.
  3. The same URL string is passed to os.makedirs() as though it were a local directory.
  4. Transcript storage fails or creates an unintended path based on the current working directory and supplied string.
  5. The resulting transcript may be unavailable, misplaced, or written somewhere the operator did not expect.

Impact Assessment

The primary impact is loss of availability and integrity in transcript storage. Processing ...[truncated 342 chars]

Remediation
View remediation

Remediation Suggestions

  1. Use a proper argument parser with separate named options such as --audio-url and --output-dir.
  2. Keep the output directory default independent of the URL argument.
  3. Validate the remote URL's scheme and hostname against an allowlist.
  4. Normalize and validate the output path before creating directories.
  5. Restrict transcript output to an approved base directory and reject path traversal outside it.
  6. Perform local destination validation before submitting the external transcription task.
  7. Add tests covering default output, explicit output directories, malformed URLs, and URL-like path strings.
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Taint TrackingDirect Taint Flow, Variable-Mediated Taint Flow, Credential Exfiltration Chain
  • MCP Least PrivilegeUnderdeclared Capability, Wildcard Permission, Missing Permission Declaration
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
Findings (16)

Tainted flow: 'req' from os.getenv (line 29, credential/environment) → urllib.request.urlopen (network output)

Critical
Category
Data Flow
Confidence
90% confidence
Finding

Credentials or environment variables flow to a network sink. This is a high-confidence indicator of credential exfiltration.

Content

Scanner excerpt · scripts/transcribe_meixin.py (reported line 22)May include surrounding context.

python
req = urllib.request.Request(url, data=json.dumps(payload, ensure_ascii=False).encode(),
        headers={"Authorization": f"Bearer {API_KEY}", "Content-Type": "application/json",
                 "X-DashScope-Async": "enable"}, method="POST")
    with urllib.request.urlopen(req, timeout=10) as r:
        return json.loads(r.read())["output"]["task_id"]

def poll_result(task_id: str) -> dict:

Tainted flow: 'req' from os.getenv (line 29, credential/environment) → urllib.request.urlopen (network output)

Critical
Category
Data Flow
Confidence
90% confidence
Finding

Credentials or environment variables flow to a network sink. This is a high-confidence indicator of credential exfiltration.

Content

Scanner excerpt · scripts/transcribe_meixin.py (reported line 30)May include surrounding context.

python
req = urllib.request.Request(url, data=json.dumps(payload, ensure_ascii=False).encode(),
        headers={"Authorization": f"Bearer {API_KEY}", "Content-Type": "application/json",
                 "X-DashScope-Async": "enable"}, method="POST")
    with urllib.request.urlopen(req, timeout=10) as r:
        return json.loads(r.read())["output"]["task_id"]

def poll_result(task_id: str) -> dict:

Tp4

High
Category
MCP Tool Poisoning
Confidence
96% confidence
Finding

The skill claims to accept meeting audio and generate structured minutes, but the actual documented implementation mainly submits audio to an external ASR service and polls results, while omitting the advertised downstream processing and hiding important operational details such as reliance on public URLs and external transmission. This mismatch is dangerous because users may provide sensitive internal recordings under incomplete or misleading expectations about where data goes and what processing occurs.

Content

No source excerpt is available for this finding.

Missing User Warnings

High
Category
Not specified by scanner
Confidence
96% confidence
Finding

The skill processes internal meeting recordings but does not clearly warn users that audio will be transmitted to an external ASR provider. Because meeting recordings may contain confidential business, personal, or regulated information, lack of explicit notice and consent creates significant privacy, compliance, and data-governance risk.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

High
Category
Not specified by scanner
Confidence
99% confidence
Finding

The documentation contains a hard-coded DashScope API key fallback directly in the sample code. Exposed credentials can be abused by anyone who reads the file, leading to unauthorized use of the external service, billing abuse, and possible attribution of malicious activity to the organization.

Content

No source excerpt is available for this finding.

Missing User Warnings

High
Category
Not specified by scanner
Confidence
99% confidence
Finding

Displaying a hard-coded API key in markdown without a warning about credential handling exposes a live secret and normalizes unsafe secret-management practices. Anyone with access to the skill can reuse the credential or propagate the insecure pattern elsewhere.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

High
Category
Not specified by scanner
Confidence
100% confidence
Finding

The script embeds a live-looking DashScope API key directly in source as a fallback value. Anyone with access to the skill can reuse the credential for unauthorized API calls, incur cost, and potentially access or submit sensitive meeting transcription workloads under the organization's account.

Content

No source excerpt is available for this finding.

Missing User Warnings

High
Category
Not specified by scanner
Confidence
99% confidence
Finding

Using a hardcoded fallback API key for outbound requests means the script will silently operate with an embedded secret even when the operator has not configured credentials. This hides credential-handling risk, encourages secret reuse in code distribution, and exposes the account to abuse if the script is shared or leaked.

Content

No source excerpt is available for this finding.

Ssd 3

High
Category
Not specified by scanner
Confidence
99% confidence
Finding

The script instructs operators to provide a publicly accessible CDN URL for meeting audio, effectively encouraging exposure of sensitive recordings to the internet before transcription. For internal health-related meeting content, this creates a substantial risk of unauthorized access, indexing, link leakage, or long-lived public retention outside controlled channels.

Content

No source excerpt is available for this finding.

Undeclared Tool Scope

Medium
Category
MCP Least Privilege
Confidence
77% confidence
Finding

The skill declares broad optional tools and demonstrates capabilities to read environment variables, make outbound network requests, and persist outputs, but it does not define an explicit tool scope such as permissions or allowed-tools. In a skill that handles sensitive meeting audio and internal memory files, unclear scoping increases the chance of unintended data access or exfiltration and reduces enforceability of least privilege.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
83% confidence
Finding

The sample request hardcodes language_hints to zh, and the document otherwise presents the workflow as automatic rather than opt-in for language selection. This can violate language/locale policy when the skill is used in mixed-language contexts because it imposes a specific language without presenting a user choice or explicit documented limitation.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
90% confidence
Finding

The skill writes generated minutes and raw transcripts into persistent workspace storage without clearly warning users. Persistent storage of complete transcripts from internal meetings increases the risk of unauthorized later access, over-retention of sensitive information, and accidental reuse in unrelated contexts.

Content

No source excerpt is available for this finding.

Ssd 3

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

The instructions emphasize full capture and retention of all meeting speech and user-provided audio content in persistent memory. In the context of internal corporate meetings, comprehensive retention materially raises privacy, confidentiality, insider-risk, and compliance exposure, especially when paired with external transcription and knowledge-base linkage.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

The script uploads meeting audio for third-party transcription without any explicit consent, privacy notice, or warning that potentially sensitive internal meeting content is being sent off-platform. In the context of health-insurance-related internal meetings, this is more dangerous because the audio may contain confidential business data or regulated personal information.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
81% confidence
Finding

The request payload hardcodes language_hints to zh, and the surrounding user-facing text is exclusively Chinese, indicating a fixed language/locale behavior. There is no opt-in, language selection, or documented justification that this tool is intentionally limited to Chinese-only use.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Low
Category
Not specified by scanner
Confidence
94% confidence
Finding

The core capability section states the output is structured meeting minutes '不含表格', and the quality standard later repeats '不使用表格'. However, the actual output template includes a markdown table in the appendix for key term explanations, which directly contradicts the stated formatting rule.

Content

No source excerpt is available for this finding.