T01 · Skill Instruction Hijacking
- Location
SKILL.md:21- Finding
Skill instructions override normal agent routing and treat untrusted transcripts as user commands
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
The skill performs legitimate cloud transcription work, but it also claims unsafe instruction priority, can make privileged system changes, and can send audio or credentials to environment-controlled endpoints.
Review this skill before installing. Use it only if you are comfortable sending audio to Volcengine/Bytedance services, avoid sensitive recordings unless you have the right approvals, preinstall ffmpeg yourself instead of allowing autonomous package installation, set only trusted HTTPS endpoints, and avoid providing ARK_SKILL_API_KEY unless you intentionally want the skill to manage and persist speech API keys.
SKILL.md:21Skill instructions override normal agent routing and treat untrusted transcripts as user commands
scripts/api_key.py:94ARK bearer credential can be transmitted to an arbitrary environment-controlled endpoint
scripts/asr_flash.py:39ASR credentials and private audio can be uploaded to arbitrary configured endpoints
scripts/ensure_ffmpeg.py:52Automatic ffmpeg installation performs privileged host-wide package and repository changes
scripts/api_key.py:64Retrieved speech API keys are persisted to a dotenv file without safe file creation
Credentials or environment variables flow to a network sink. This is a high-confidence indicator of credential exfiltration.
}
log(f"正在从飞书下载语音文件: file_key={file_key}")
resp = requests.get(download_url, headers=headers, timeout=30)
if resp.status_code != 200:
fail(
The ASR destination is fully overrideable via the MODEL_SPEECH_ASR_API_BASE environment variable, and the request includes authentication headers plus either raw audio content or a supplied audio URL. If an attacker can influence environment variables or runtime configuration, they can redirect requests to an arbitrary endpoint and capture API keys and sensitive audio data.
}
log("正在调用火山引擎 ASR...")
resp = requests.post(ASR_ENDPOINT, json=payload, headers=headers, timeout=60)
status_code = resp.headers.get("X-Api-Status-Code", "")
if status_code != "20000000":
Credentials or environment variables flow to a network sink. This is a high-confidence indicator of credential exfiltration.
log(f"正在提交标准版识别任务: task_id={task_id}")
try:
resp = requests.post(
ASR_STANDARD_SUBMIT_URL,
data=json.dumps(request_body),
headers=headers,
Credentials or environment variables flow to a network sink. This is a high-confidence indicator of credential exfiltration.
headers["X-Tt-Logid"] = x_tt_logid
try:
resp = requests.post(
ASR_STANDARD_QUERY_URL,
data=json.dumps({}),
headers=headers,
The skill is described as an ASR/transcription tool, but it also instructs automatic system-level installation of ffmpeg/ffprobe, including potential sudo-based package installation and repository enablement. That is materially broader than simple transcription and can modify the host system in security-sensitive ways without a clear consent gate.
The skill is described as an ASR/transcription tool, but it also instructs automatic system-level installation of ffmpeg/ffprobe, including potential sudo-based package installation and repository enablement. That is materially broader than simple transcription and can modify the host system in security-sensitive ways without a clear consent gate.
The skill is described as an ASR/transcription tool, but it also instructs automatic system-level installation of ffmpeg/ffprobe, including potential sudo-based package installation and repository enablement. That is materially broader than simple transcription and can modify the host system in security-sensitive ways without a clear consent gate.
Code accesses credential files (SSH keys, AWS credentials, etc.). This could indicate credential theft attempts.
def _dotenv_path() -> str:
return os.path.join(os.path.dirname(os.path.abspath(__file__)), ".env")
def _load_dotenv_if_available(dotenv_path: str) -> None:
The script explicitly seeks privilege escalation paths via sudo/root detection so it can perform package installation automatically. For a voice-to-text skill, autonomous privileged system administration is out of scope and dangerous because it can modify the host with elevated permissions, increasing blast radius if the skill is misused or chained with other weaknesses.
The skill declares shell, network, and environment-variable dependent behavior but does not constrain or document tool scope via explicit permissions/allowed-tools. That increases the chance an agent executes broader-than-expected commands, especially because the skill instructs running Python scripts, accessing external services, and modifying the local environment.
The skill sends user audio or audio URLs to an external ASR provider but does not prominently warn that potentially sensitive voice content will leave the local environment. This can lead to unintentional disclosure of personal, confidential, or regulated data, especially because the skill is triggered automatically for audio inputs.
The skill directs the agent to automatically install ffmpeg/ffprobe when missing, but it does not provide a clear warning that this will change the local system and may require elevated privileges. Silent or implicit system modification is risky in agentic contexts because users may not expect package installations or repository changes.
The skill contains logic to enumerate existing remote API keys and create new ones using a broader privileged token, which is beyond the narrow scope of speech-to-text transcription. This expands the skill from consuming a configured credential to provisioning credentials, increasing blast radius if the skill is invoked unexpectedly or abused.
The skill persists a newly discovered or created API key to a local .env file, creating long-lived plaintext credential storage on disk. Even with a best-effort chmod, this can expose secrets to other local users, backups, logs, packaging workflows, or accidental source inclusion.
The code automatically creates a service API key and stores it for future use without an explicit user-facing consent step. In a voice-to-text skill, silent credential provisioning is unnecessary for core functionality and can enable unintended account changes or persistence of access beyond the original session.
This code file contains natural-language documentation in Chinese and continues with Chinese-only user-facing messages throughout the CLI flow. The policy specifically flags language or locale constraints when a skill forces a specific language without user opt-in, and this script does not offer an alternative locale for its interface text.
The script downloads audio from Feishu and then sends audio data or audio URLs to the Volcengine ASR service, which is a network operation involving potentially sensitive user data. Although there are operational log messages, there is no clear user-facing warning or explanatory comment/docstring disclosing that audio content will be sent to third-party services for processing.
Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.
}
log("正在调用火山引擎 ASR...")
resp = requests.post(ASR_ENDPOINT, json=payload, headers=headers, timeout=60)
status_code = resp.headers.get("X-Api-Status-Code", "")
if status_code != "20000000":
This code uploads user-provided audio or audio URLs to an external third-party ASR service, including full local file contents when --file is used, but does not provide an explicit warning or consent check at the point of transfer. In an agent skill context that may process voice messages automatically, this can expose sensitive speech content to an external provider without the user's clear awareness.
The module docstring is entirely in Chinese and presents usage and behavior only in that language, with no indication that the user can choose another language or locale. This can violate language/locale policy when a skill imposes a specific language by default rather than offering an opt-in or documented regional constraint.
subprocess module calls execute external commands. Without careful input validation, this enables command injection.
def run_command(command):
return subprocess.run(
command,
stdout=subprocess.PIPE,
stderr=subprocess.PIPE,
The script can enable additional repositories and install system packages, including fetching an RPM from an external repository URL, which expands system trust and changes host configuration beyond simple ASR logic. In an agent skill context, this creates unnecessary supply-chain and environment-modification risk because using the skill may trigger persistent system administration actions on the host.
Natural-language policy violations include forcing a specific language without user opt-in. The module description, usage guidance, and comments are presented in Chinese only, and the file does not indicate that this is a region-specific tool or provide an alternative language option.
subprocess module calls execute external commands. Without careful input validation, this enables command injection.
def run_command(command, timeout_seconds=None):
try:
return subprocess.run(
command,
check=True,
stdout=subprocess.PIPE,
The dependency is specified as requests>=2.28.0, which allows installation of many possible future or currently vulnerable versions and prevents reproducible builds. In a skill that processes audio from local files, URLs, and messaging platforms, a transitive install of an affected requests release could expose the skill to known client-side issues such as credential leakage or request-handling flaws.
# Voice-to-Text skill dependencies
# Ref: https://www.volcengine.com/docs/6561/1354870
requests>=2.28.0