Back to skill

Security audit

Byted Kickart Marketing Material Generator

Security checks for vulnerabilities and agentic risk

Overview

The skill has a plausible marketing-video purpose, but it handles cloud credentials, private media, logs, and background execution in ways that need careful review before use.

Review this skill before installing. Use only tightly scoped, revocable Volcengine credentials, do not paste long-lived secrets into chat, avoid running it where ARK_SKILL_API_BASE can be altered by untrusted files or users, and assume logs or /tmp state may contain sensitive task, media, and credential data until the package is fixed.

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 (5)

T09 · Insecure Skill Coding Practices

Error
Location
references/火山鉴权指南.md:11
Finding

Authentication credentials are exposed through chat and terminal output

Content
View full analysis
export SECRET_ACCESS_KEY= ``` ### Technical Analysis The authentication guide instructs the agent to ask users to submit long-lived Volcengine access credentials directly through chat. It then requires an environment-check command that prints the complete bearer token, Access Key ID, and Secret Access Key. Although later instructions state that credentials must not be disclosed, the mandatory command directly contradicts that requirement. The values can be captured in: - Chat transcripts and message retention systems - Agent execution traces - Terminal output - Shell audit records or command history - Debugging and support bundles - Monitoring systems that collect command output An Access Key ID and Secret Access Key are account-level authentication material. Their exposure is not necessary to determine whether authentication is configured; checking whether the variables are set would be sufficient. ### Attack Path 1. The Skill determines that authentication is not configured. 2. Following the guide, it asks the user to send an Access Key ID and Secret Access Key in chat. 3. The credentials become part of the retained conversation history. 4. The Skill exports the credentials into its environment. 5. The mandatory environment check prints the complete credentials. 6. An attacker with access to conversation records, executi ...[truncated 605 chars]
Remediation
View remediation

T09 · Insecure Skill Coding Practices

Error
Location
scripts/servicev2.py:49
Finding

Bearer authentication tokens and sensitive request bodies are persisted in plaintext logs

Content
View full analysis
>> {method.upper()} {url} {headers} {body}") response = requests.request( method=method.upper(), url=url, headers=headers, data=body, timeout=30 ) logging.info(f"<<< {response.headers} {response.text}") ``` The project-wide log configuration is: ```python filename = f"/tmp/openclaw/kickart/info.{time.strftime('%Y%m%d', time.localtime())}.log" logging.basicConfig( level=logging.INFO, filename=filename, format="%(asctime)s - %(levelname)s - %(message)s", datefmt="%Y-%m-%d %H:%M:%S", ) ``` ### Technical Analysis The request logger serializes the complete `headers` dictionary. That dictionary includes the unredacted `Authorization: Bearer ...` header. It also serializes the complete request body and response text. The request body may contain product links, prompts, uploaded-media URLs, storyboards, publishing content, task parameters, or other user-provided marketing data. Responses may contain task identifiers and generated-content URLs. The logs are written under a predictable `/tmp/openclaw/kickart/` path. The code does not explicitly create the log file with restrictive permissions, enforce retention, rotate credentials out of logs, or redact sensitive fields. ### Attack Path 1. The user invokes a task submission or query through the bearer-token authentication mode. 2. `_do_request` builds an authorization header containing `ARK_SKILL_API_KEY`. 3. The full headers and request body are written to the plaintext log. 4. The response headers and response body are also persisted. 5. A local user, compromised process, support tool, backup collec ...[truncated 643 chars]
Remediation
View remediation
>> %s %s headers=%s", method.upper(), url, safe_headers) ``` 3. Do not log complete request or response bodies at `INFO` level. 4. If diagnostic body logging is essential, use an opt-in debug mode and redact media URLs, prompts, user identifiers, and task identifiers. 5. Create the log directory and files with restrictive permissions such as `0700` for the directory and `0600` for files. 6. Implement rotation, short retention, and secure deletion. 7. Rotate any bearer token that may already have been written to these logs. 8. Review other logging statements such as command-line logging because command arguments may also contain user metadata or sensitive URLs. ]]>

T09 · Insecure Skill Coding Practices

Error
Location
scripts/chunks.py:79
Finding

Environment-controlled API destination can receive bearer tokens, task data, and uploaded media without validation

Content
View full analysis
dict: headers = { "ServiceName": service, "Authorization": f"Bearer {arkclaw_token}", "Content-Type": "application/octet-stream", } url = f"{host}/?Action={action}&Version={VERSION}" if extra_query: url += "&" + urlencode(extra_query) debug_print(f"url is {url}, data size is {len(data)}") resp = requests.post(url, data=data, headers=headers, timeout=60) ``` The same design is used for task requests: ```python ADDR = os.getenv("ARK_SKILL_API_BASE") TOKEN = os.getenv("ARK_SKILL_API_KEY") or "" url = f"{ADDR}?{query_string}" headers = { "Content-Type": "application/json", "ServiceName": service, "Authorization": "Bearer " + TOKEN, } response = requests.request( method=method.upper(), url=url, headers=headers, data=body, timeout=30 ) ``` ### Technical Analysis Uploading user-selected media to a remote service is necessary for the declared functionality. However, the implementation does not constrain `ARK_SKILL_API_BASE` to a trusted HTTPS origin. There is no validation of: - URL scheme - Expected Volcengine hostname - Port - Embedded credentials - IP-literal destinations - Loopback, private, or link-local addresses - Trusted certificate identity beyond the behavior of the supplied URL Consequently, anyone able to influence the process environment can redirect credential-bearing requests ...[truncated 1378 chars]
Remediation
View remediation

T09 · Insecure Skill Coding Practices

Warning
Location
scripts/media.py:92
Finding

Unsanitized session identifiers allow path traversal in local state operations

Content
View full analysis
pd.DataFrame: path = Path(f"/tmp/kickart/material_state_{session_id}.csv") if not path.exists(): return pd.DataFrame(columns=COLUMNS) return pd.read_csv(path, header=None, names=COLUMNS) def save(session_id: str, df: pd.DataFrame): path = Path(f"/tmp/kickart/material_state_{session_id}.csv") os.makedirs(path.parent, exist_ok=True) df.to_csv(path, index=False, header=False) def remove(session_id: str): path = Path(f"/tmp/kickart/material_state_{session_id}.csv") os.remove(path) ``` The value is accepted directly from the command line: ```python @click.option("--session-id", "-s", required=True, help="Session ID") def add(file, session_id): ... df = load(session_id) df.loc[len(df)] = row save(session_id, df) ``` ```python @click.option("--session-id", "-s", required=True, help="Session ID") def clear(session_id): remove(session_id) ``` ### Technical Analysis The session identifier is interpolated directly into a filesystem path without character validation, canonicalization, or a containment check. A value containing path separators and `..` components can resolve outside `/tmp/kickart`. The fixed `material_state_` prefix and `.csv` suffix constrain the exact target name, but they do not guarantee that the resolved path remains in the intended directory. Depending on the supplied identifier and filesystem layout, the `load`, `save`, and `clear` operations can access unintended CSV-suffixed paths under the process account's permissions. The `/tmp` location also creates symlink and multi-user isolation concerns because the code does not use exclusive file creation or verify that the target is a regular file owned by the current process. ### Attack Path ...[truncated 981 chars]
Remediation
View remediation
str: if not SESSION_RE.fullmatch(session_id): raise ValueError("Invalid session identifier") return session_id ``` 2. Resolve and verify containment before every operation: ```python STATE_ROOT = Path("/tmp/kickart").resolve() def state_path(session_id: str) -> Path: validate_session_id(session_id) candidate = (STATE_ROOT / f"material_state_{session_id}.csv").resolve() if candidate.parent != STATE_ROOT: raise ValueError("State path escapes the state directory") return candidate ``` 3. Create the state directory with restrictive permissions. 4. Use atomic writes through a securely created temporary file followed by `os.replace`. 5. Reject symlinks and verify file ownership before reading or deleting state. 6. Catch `FileNotFoundError` in `remove` so clearing an absent session does not crash. 7. Prefer an internal keyed state database rather than deriving filenames from externally supplied identifiers. ]]>

T09 · Insecure Skill Coding Practices

Warning
Location
scripts/material.py:12
Finding

Redirects from approved product domains can reach untrusted or internal network destinations

Content
View full analysis
Result: try: # Validate domain parsed_original = urlparse(url) original_domain = parsed_original.netloc allowed_domains = ["haohuo.jinritemai.com", "v.douyin.com"] if original_domain not in allowed_domains: return Result( code="-1", message=f"URL domain is unsupported; only the following domains are supported: {', '.join(allowed_domains)}", ) response = requests.head(url, allow_redirects=True) parsed = urlparse(response.url) query = {k: v for k, v in parse_qs(parsed.query).items() if k in keys} simplified_url = parsed._replace(query=urlencode(query, doseq=True)).geturl() return Result(code="0", message=simplified_url) ``` ### Technical Analysis The function validates only the hostname in the original URL. It then enables automatic redirect following without validating any redirect destination. An approved short-link service is expected to redirect, but each redirect target must be treated as a new network destination. If an approved URL can be made to redirect to an arbitrary location, or if the trusted service is compromised, the Skill may issue a server-side HEAD request to: - Loopback services - Private network addresses - Link-local services - Cloud metadata endpoints - Internal administrative interfaces - Attacker-controlled hosts The request does not carry the Skill's API authorization header, which limits credential impact, but it still exposes the network position and reachability of the machine running the Skill. ### Attack Path 1. An attacker supplies a syntactically approved URL under `v.douyin.com` or `haohuo.jinritemai.com`. 2. That URL responds with ...[truncated 935 chars]
Remediation
View remediation
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • Taint TrackingDirect Taint Flow, Variable-Mediated Taint Flow, Credential Exfiltration Chain
  • MCP Least PrivilegeUnderdeclared Capability, Wildcard Permission, Missing Permission Declaration
Findings (70)

Missing User Warnings

Critical
Category
Not specified by scanner
Confidence
100% confidence
Finding

The guide tells users to send Access Key ID and Secret Access Key directly in chat, which is one of the least safe channels for transmitting long-lived cloud credentials. In the context of a marketing-content skill, this is especially dangerous because the capability does not justify direct handling of privileged infrastructure secrets, creating severe risk of account compromise if the credentials are exposed or retained.

Content

No source excerpt is available for this finding.

Tainted flow: 'headers' from os.getenv (line 349, credential/environment) → requests.post (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/chunks.py (reported line 213)May include surrounding context.

python
debug_print(f"canonical_string: {repr(canonical_string)}")
    debug_print(f"sign_string: {repr(sign_string)}")

    resp = requests.post(url, data=body_bytes, headers=headers, timeout=30)

    try:
        result = resp.json()

Tainted flow: 'headers' from os.getenv (line 349, credential/environment) → requests.post (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/chunks.py (reported line 248)May include surrounding context.

python
debug_print(f"canonical_string: {repr(canonical_string)}")
    debug_print(f"sign_string: {repr(sign_string)}")

    resp = requests.post(url, data=body_bytes, headers=headers, timeout=30)

    try:
        result = resp.json()

Tainted flow: 'headers' from os.getenv (line 349, credential/environment) → requests.post (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/chunks.py (reported line 280)May include surrounding context.

python
if extra_query:
        url += "&" + urlencode(extra_query)
    debug_print(f"url is {url}, data size is {len(data)}")
    resp = requests.post(url, data=data, headers=headers, timeout=60)

    try:
        result = resp.json()

Tainted flow: 'headers' from os.getenv (line 349, credential/environment) → requests.post (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/chunks.py (reported line 359)May include surrounding context.

python
if extra_query:
        url += "&" + urlencode(extra_query)
    debug_print(f"url is {url}, data size is {len(data)}")
    resp = requests.post(url, data=data, headers=headers, timeout=60)

    try:
        result = resp.json()

Tp4

High
Category
MCP Tool Poisoning
Confidence
96% confidence
Finding

This finding suggests the entry flow mainly performs task-result querying and uses fixed test assets/templates rather than the full user-driven generation workflow claimed in the description. That is risky because it misrepresents both capability and data handling, and fixed test resources can accidentally mix tenant data or produce confusing, unauthorized outputs.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
95% confidence
Finding

This finding suggests the entry flow mainly performs task-result querying and uses fixed test assets/templates rather than the full user-driven generation workflow claimed in the description. That is risky because it misrepresents both capability and data handling, and fixed test resources can accidentally mix tenant data or produce confusing, unauthorized outputs.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding

This finding suggests the entry flow mainly performs task-result querying and uses fixed test assets/templates rather than the full user-driven generation workflow claimed in the description. That is risky because it misrepresents both capability and data handling, and fixed test resources can accidentally mix tenant data or produce confusing, unauthorized outputs.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
96% confidence
Finding

This finding suggests the entry flow mainly performs task-result querying and uses fixed test assets/templates rather than the full user-driven generation workflow claimed in the description. That is risky because it misrepresents both capability and data handling, and fixed test resources can accidentally mix tenant data or produce confusing, unauthorized outputs.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

This finding suggests the entry flow mainly performs task-result querying and uses fixed test assets/templates rather than the full user-driven generation workflow claimed in the description. That is risky because it misrepresents both capability and data handling, and fixed test resources can accidentally mix tenant data or produce confusing, unauthorized outputs.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
95% confidence
Finding

This finding suggests the entry flow mainly performs task-result querying and uses fixed test assets/templates rather than the full user-driven generation workflow claimed in the description. That is risky because it misrepresents both capability and data handling, and fixed test resources can accidentally mix tenant data or produce confusing, unauthorized outputs.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
94% confidence
Finding

This finding suggests the entry flow mainly performs task-result querying and uses fixed test assets/templates rather than the full user-driven generation workflow claimed in the description. That is risky because it misrepresents both capability and data handling, and fixed test resources can accidentally mix tenant data or produce confusing, unauthorized outputs.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
96% confidence
Finding

This finding suggests the entry flow mainly performs task-result querying and uses fixed test assets/templates rather than the full user-driven generation workflow claimed in the description. That is risky because it misrepresents both capability and data handling, and fixed test resources can accidentally mix tenant data or produce confusing, unauthorized outputs.

Content

No source excerpt is available for this finding.

Missing User Warnings

High
Category
Not specified by scanner
Confidence
99% confidence
Finding

The command explicitly prints full values of ARK_SKILL_API_KEY, ACCESS_KEY_ID, and SECRET_ACCESS_KEY to output. This can expose secrets to chat history, terminal logs, observability tooling, screenshots, or any downstream system that captures command output.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

High
Category
Not specified by scanner
Confidence
99% confidence
Finding

The guide instructs the agent to read sensitive credential environment variables by echoing them in full and to solicit cloud access keys directly from the user in chat. For a marketing-video generation skill, collecting raw cloud credentials is unnecessary and materially increases the chance of credential exposure, reuse, or misuse.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

High
Category
Not specified by scanner
Confidence
99% confidence
Finding

The document is internally contradictory: it claims secrets must not be exposed, but earlier steps require echoing complete secret values and asking the user to send them in plaintext. This inconsistency creates a false sense of safety while normalizing unsafe secret-handling practices that can leak credentials through logs, transcripts, and operator visibility.

Content

No source excerpt is available for this finding.

Ssd 3

High
Category
Not specified by scanner
Confidence
99% confidence
Finding

The workflow instructs the agent to accept cloud credentials pasted into chat and then export them into the session environment for subsequent use. Even if intended to be temporary, this creates unnecessary exposure in message history and runtime state, and it expands the blast radius if the agent, logs, or connected tools are compromised.

Content

No source excerpt is available for this finding.

Missing User Warnings

High
Category
Not specified by scanner
Confidence
99% confidence
Finding

When debug logging is enabled, the code prints full request headers, including Authorization material, and request bodies or signing context. This can leak bearer tokens, derived authentication data, and possibly uploaded content metadata into logs, enabling credential theft or replay by anyone with log access.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
Not specified by scanner
Confidence
98% confidence
Finding

The implemented behavior performs Ark Claw package query/registration, which does not match the declared purpose of generating and publishing marketing videos. Capability drift like this is dangerous because a user or calling platform may invoke the skill expecting media-generation actions while the code instead performs account-affecting service registration against a backend API.

Content

No source excerpt is available for this finding.

Undeclared Tool Scope

Medium
Category
MCP Least Privilege
Confidence
92% confidence
Finding

The skill text instructs the agent to use high-risk capabilities including environment-variable access, local file reads/writes, shell execution, and network calls, but it declares no explicit tool scope or permission boundary. That creates an over-privileged execution surface where a reviewer or platform cannot easily constrain what the skill may access, increasing the chance of secret exposure or unintended system modification.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

The skill mandates clearing uploaded material records immediately after delivering a generated video, but the description does not warn users about this destructive action or obtain confirmation. Silent deletion is dangerous because it can cause data loss, break auditability/reproducibility, and prevent users from reusing or verifying assets they expected to remain available.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The guide defines trigger phrases such as '查看结果', '生成怎么样了', and '好了吗', which are broad conversational phrases that can easily appear in unrelated user messages. In an agent skill that can invoke task polling and read back task result files, overly broad triggers increase the chance of unintended tool execution, context confusion, and disclosure of the wrong task's status or output.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The mandated agent workflow specifies exact Chinese phrases the agent must use with users, such as the required prompts and status messages, and does so as a blanket requirement. This imposes a specific language on all interactions without documenting user choice, opt-in, or a region-specific justification.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The guide requires persisting granular user requirements and task metadata in session context, including platform, target audience, selling points, task ID, submission time, and parameters. For a creative-analysis skill, retaining this level of profiling and operational metadata beyond immediate execution expands data exposure and creates unnecessary privacy and misuse risk if session data is later accessed, reused, or leaked.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

The guide mandates executing a clear command that removes current session material records immediately after user confirmation, which introduces destructive behavior into the agent workflow. Even if described as clearing only local session associations, forced deletion can surprise users, interfere with auditability/recovery, and be abused to erase evidence of what materials were used.

Content

No source excerpt is available for this finding.