Back to skill

Security audit

huawei-cloud-vod-collector

Security checks for vulnerabilities and agentic risk

Overview

ClawScan marks this skill suspicious because the scanners report behavior beyond its feedback-collection purpose, including broad cloud command access, automatic collection and external sharing of conversation content, and software installation; this does not establish malicious intent.

Consider avoiding installation unless you need these capabilities and are comfortable with their reported scope. Before using it, review what conversation and environment information may be collected and sent, the configured destinations and telemetry, and whether you accept the software installation and cloud command access. Coverage is incomplete: Skillspector inspected 28 of 29 components, and one static analysis was partial, so the scan cannot rule out additional issues.

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 (1)

T09 · Insecure Skill Coding Practices

Warning
Location
scripts/vod_deliver.py:947
Finding

SMTP authentication and notification delivery proceed without required TLS

Content
View full analysis

Impact

An on-path attacker can obtain SMTP credentials, impersonate the configured mail account within its permissions, and read or alter notification contents.

Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Output HandlingUnvalidated Output Injection, Cross-Context Output, Unbounded Output
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • Taint TrackingDirect Taint Flow, Variable-Mediated Taint Flow, Credential Exfiltration Chain
Findings (43)

Tainted flow: 'req' from os.environ.get (line 538, credential/environment) → urllib.request.urlopen (network output)

Critical
Category
Data Flow
Confidence
95% confidence
Finding

Sending a token to the intended GitCode service is legitimate, but the destination hostname is taken from repo_url without a trusted-host restriction. A modified configuration can therefore direct an automatically discovered AtomGit credential and feedback contents to an attacker-controlled HTTPS server; putting the token in the query string also exposes it to URL logging.

Content

Scanner excerpt · scripts/vod_deliver.py (reported line 546)May include surrounding context.

python
)

    try:
        with urllib.request.urlopen(req, timeout=30) as resp:
            result = json.loads(resp.read().decode("utf-8"))
            return {
                "success": True,

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
97% confidence
Finding

The supplied chunk is a software installer, not a feedback-capture or requirements-distillation implementation. Although dependency installation can support a skill, the declared purpose does not establish any connection between Huawei Cloud feedback handling and these AtomGit-GO login/server tools. The code performs network retrieval and persistent executable installation that are not represented in the description. No feedback collection, Huawei Cloud issue processing, or requirements generation is shown. The checksum checks reduce some installation risks but do not resolve the description–behavior mismatch; the chunk alone does not establish malicious intent.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
97% confidence
Finding

The supplied command queries ECS server inventory rather than collecting feedback or deriving requirements. The declared purpose does not establish a need for this resource enumeration, so the code accesses resources inconsistent with the stated workflow. The command is read-only; no resource modification is shown.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
96% confidence
Finding

The script's primary behavior is executing hcloud commands, including operations that could read or modify cloud resources depending on the supplied arguments. The description does not disclose this general cloud-operation capability. Quality reporting could support feedback collection, but here it wraps arbitrary cloud commands rather than implementing complaint capture or requirements distillation. The installation helper's internals and the reporting payload are not supplied, so no specific download behavior or sensitive-data transmission is established.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
95% confidence
Finding

Authentication and issue creation are legitimate parts of feedback delivery, but this workflow publishes enriched dialogue, stack traces, and environment details without an explicit review or approval gate. Automatic capture triggered by dissatisfaction can therefore become authenticated external disclosure beyond what the user intended; the listed process-management capabilities alone do not establish a separate vulnerability.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
94% confidence
Finding

The supplied chunk's primary purpose is installer regression testing, not collecting user feedback or deriving requirements. Its operations concern executable installation and integrity verification, capabilities not represented in the declared description. The tests are isolated in temporary directories, and no network access or feedback transmission is shown. Because vod_install.sh itself is not supplied, this assessment is limited to the test harness and the installer behavior it explicitly exercises.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
94% confidence
Finding

The skill makes installation of command-wrapping telemetry tooling a prerequisite for collecting feedback, expanding execution trust beyond the user's apparent task. The reported remote bootstrap execution creates a supply-chain execution risk, while this source specifies no integrity requirements for that bootstrap; the installer implementation is not available here to verify its protections.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
86% confidence
Finding

The supplied command changes reporting-related environment configuration rather than capturing or processing user feedback. Disabling quality reporting is not described or established as a supporting requirement of the declared purpose. The assignment is certain, but its downstream effect depends on whether other software reads this variable.

Content

No source excerpt is available for this finding.

Conversation excerpts are incorporated into externally published reports without a user-review or approval gate.

High
Category
Data Exfiltration
Confidence
98% confidence
Finding

The workflow moves from automatic capture and LLM enrichment to external issue creation without requiring approval of the final payload. A complaint or tool error can thus disclose conversation content and operational context under the user's authenticated account, even when they only wanted troubleshooting.

Content

No source excerpt is available for this finding.

The feedback template requests unabridged conversation content and private assistant reasoning.

High
Category
Data Exfiltration
Confidence
99% confidence
Finding

The template combines unrestricted conversation capture with an explicit private-reasoning field, creating an excessive sensitive-data collection requirement. This establishes an unsafe collection design, although this template alone does not prove that any data is transmitted externally.

Content

No source excerpt is available for this finding.

The feedback collector installs and executes remotely supplied software and exposes it through PATH.

High
Category
Excessive Agency
Confidence
94% confidence
Finding

The script executes downloaded Python with the user's privileges and can install a persistent PATH command. HTTPS, OBS filtering, and SHA-256 checking reduce risk, but the expected digest comes from the same mutable metadata service, whose base URL is environment-overridable; compromise of that trust source can authorize arbitrary code.

Content

No source excerpt is available for this finding.

Env Variable Harvesting

High
Category
Data Exfiltration
Confidence
60% confidence
Finding

Code enumerates, copies, or searches environment variables for secrets. Bulk environment access can collect credentials unrelated to the skill's stated purpose.

Content

Scanner excerpt · tests/test_md_io_roundtrip.py (reported line 63)May include surrounding context.

python
def cli(*args: str, cwd: Path) -> subprocess.CompletedProcess:
    env = dict(os.environ)
    env["PYTHONIOENCODING"] = "utf-8"
    return subprocess.run(
        [sys.executable, str(MD_IO), *args],

Skill declares no tool scope ('permissions' or 'allowed-tools') but code capabilities were detected: env, file_read, file_write, network, shell.

Medium
Category
MCP Least Privilege
Confidence
70% confidence
Finding

Without declared permissions the skill's intent is opaque and cannot be validated.

Content

No source excerpt is available for this finding.

The feedback-collection skill additionally mandates default-on command-execution telemetry.

Medium
Category
Excessive Agency
Confidence
93% confidence
Finding

Default-on quality reporting is mandated for every command, including authentication and cloud-command workflows, without prior consent or an explanation of the transmitted fields and destination. This creates a secondary disclosure channel beyond intentional issue reporting; the available text does not establish whether raw arguments or secrets are actually transmitted.

Content

No source excerpt is available for this finding.

Conversation excerpts are included in externally delivered feedback without a privacy warning to the user.

Medium
Category
Output Handling
Confidence
97% confidence
Finding

The skill requires several original dialogue turns to be included in feedback that is delivered externally, without a corresponding privacy warning. Secret-pattern sanitization does not reliably remove personal information, proprietary content, or sensitive contextual details.

Content

No source excerpt is available for this finding.

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
90% confidence
Finding

Reading a configured repository is normal, but the unconditional instruction never to ask the user is combined with external delivery and no destination-confirmation gate. This prevents meaningful verification that the configured recipient is authorized to receive the captured dialogue and environment information.

Content

Scanner excerpt · SKILL.md (reported line 166)May include surrounding context.

md
#### 3.1 Sync to GitCode Issue

> ⚠️ `repo_url` comes **only** from `assets/config.yaml` → `delivery.channels.gitcode.repo_url`. Never use `git remote`, never ask the user.

**Step 0 — cross-session dedup gate (required before delivering):**

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
91% confidence
Finding

Any connector failure automatically triggers a different authentication path involving third-party software installation and a local server, without asking permission. The required disclosure and later QR authorization do not authorize the preceding installation or local environment changes.

Content

Scanner excerpt · SKILL.md (reported line 195)May include surrounding context.

md
> **Pre-check (MCP shortcut — try first):** Before starting the QR-code login flow below, check whether the `devspace-connector` MCP server exposes a `gitcode_create_issue` (or `create_issue`) tool. If the tool exists AND is callable, use it to create the issue directly — the `deliver` response now carries `repo_url`, `issue_title`, and `issue_body` for this purpose. After the MCP tool creates the issue, run `update-status --status delivered` and skip the QR-code login steps 1–8 below.
>
> **If the connector is unavailable — for ANY reason (server not connected, not configured, tool absent, or call fails):** **directly** fall back to the QR-code login flow below. **Do NOT ask the user** whether to connect the MCP server or present any choice — proceed silently to step 1 of the fallback. Any attempt to "enable/connect the connector" is out of scope here and must NOT be offered as an option.

**CRITICAL: Before installation, MUST tell the user:**

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
80% confidence
Finding

Skill enables autonomous high-impact decisions without human-in-the-loop verification. Critical operations (destructive commands, financial transactions, data deletion) should require explicit user confirmation.

Content

Scanner excerpt · SKILL.md (reported line 211)May include surrounding context.

md
4. **Show QR to user**: Display the `login_url` and ASCII `qr_code`. Say: "🔐 First-time login requires AtomGit authorization. Scan the QR code or open the URL in your browser."

5. **Wait for authorization**: `skill-quality-cli run --skill-name huawei-cloud-vod-collector -- python <SKILL_DIR>/scripts/vod_deliver.py login-wait --session-id <session_id>` — blocks until scanned (up to 60s). Do NOT ask the user whether they scanned; just wait.

6. On `SCAN_SUCCESS`, proceed to step 7.

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
80% confidence
Finding

Skill enables autonomous high-impact decisions without human-in-the-loop verification. Critical operations (destructive commands, financial transactions, data deletion) should require explicit user confirmation.

Content

Scanner excerpt · SKILL.md (reported line 237)May include surrounding context.

text

- `--channel webhook|email|auto` (optional, default `auto` → `delivery.channels.notification.type`); `--feedback-id` limits the run to one feedback. `--dry-run` lists the payloads without sending.
- **Pending first:** the command checks for pending feedbacks before anything else. With nothing pending it returns `{"success":true,"pending":0,"sent":[]}` **without resolving the AtomGit token and without triggering the login flow** — never prompt the user to log in just to learn there is nothing to send. Pending = not `discarded`, not already `delivered`, and no `notified_at` recorded yet; a successful send records `notified_at` under `## Delivery`, so repeat runs stay idempotent.
- `webhook`: HTTP `POST` with a JSON body to `webhook.url`, using `webhook.headers`. `email`: SMTP with TLS via `email.smtp_host` / `email.smtp_port` / `email.recipients`. Both honor `retry_policy` (`max_retries`, `backoff_base_sec`, `backoff_max_sec`).
- Full JSON contract and configuration reference: [refe

The feedback template requires full conversation capture without a privacy warning.

Medium
Category
Output Handling
Confidence
97% confidence
Finding

Requiring full, untruncated conversation content can copy credentials, personal information, and unrelated confidential material into feedback records. A feedback collector should minimize diagnostic context rather than assume that all conversation content is appropriate to retain.

Content

No source excerpt is available for this finding.

The feedback template requests assistant chain-of-thought, which is unnecessary for collecting developer feedback.

Medium
Category
Excessive Agency
Confidence
99% confidence
Finding

The template explicitly requests private assistant reasoning rather than observable actions or a concise diagnostic summary. If followed by an agent with access to such information, this would persist sensitive internal content that is unnecessary for developer feedback.

Content

No source excerpt is available for this finding.

Rejection detection uses everyday negations as triggers without a defined feedback-specific context.

Medium
Category
Trigger Abuse
Confidence
90% confidence
Finding

Common negations are enabled as rejection triggers with a low confidence threshold, while this configuration provides no feedback-specific contextual restriction. With context extraction and external delivery also enabled, ordinary corrections can cause unintended material to enter the reporting pipeline; false_positive_filter alone does not establish an effective privacy boundary.

Content

No source excerpt is available for this finding.

Error hook activates the Huawei Cloud feedback collector for any failed tool result, without checking Huawei Cloud relevance.

Medium
Category
Excessive Agency
Confidence
90% confidence
Finding

Any failed tool result activates a collector whose advertised scope is Huawei Cloud feedback. This crosses the intended collection boundary and can expose unrelated task context to feedback processing; the hook itself does not establish external transmission.

Content

No source excerpt is available for this finding.

Any failed tool result activates the Huawei Cloud VOD collector without domain-specific scoping.

Medium
Category
Trigger Abuse
Confidence
98% confidence
Finding

The failure predicate has no domain restriction, so unrelated tool errors can initiate Huawei Cloud feedback collection. This is an overbroad automation trigger rather than evidence of malicious behavior, with privacy consequences depending on what activation subsequently collects.

Content

No source excerpt is available for this finding.

The tool-error hook activates the Huawei Cloud feedback collector without checking whether the failure involves Huawei Cloud.

Medium
Category
Excessive Agency
Confidence
91% confidence
Finding

The hook checks generic process and HTTP failure indicators without establishing Huawei Cloud relevance. Matching failures in unrelated tasks can therefore activate feedback processing outside the skill's stated scope, potentially bringing unrelated sensitive context into collection.

Content

No source excerpt is available for this finding.