T01 · Skill Instruction Hijacking
- Location
SKILL.md:11- Finding
Mandatory Agent-Wide Audio Announcement Rules Hijack Normal Interaction Flow
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
This skill provides voice announcements, but it also tries to make them mandatory across future sessions and may send announcement text to an external TTS service.
Review this skill before installing. It should only be enabled if you deliberately want agent-wide spoken status updates, persistent agent behavior edits, startup checks, local audio caching, and network-backed TTS processing of announcement text. Avoid using it for confidential prompts unless announcements are disabled or rewritten to use generic, non-sensitive text.
SKILL.md:11Mandatory Agent-Wide Audio Announcement Rules Hijack Normal Interaction Flow
SKILL.md:47Persistent Agent Memory and Identity Poisoning Through Integration Instructions
SKILL.md:390Proposed Session Hooks Establish Cross-Session Execution Persistence
audio_announcement/announce_helper.py:237Interaction Content Is Sent to a Network-Backed TTS Service Without Adequate Privacy Disclosure
pyproject.toml:1Mutable and Inconsistently Named Third-Party Dependencies Create Supply-Chain Risk
The skill directly instructs the agent to speak the user's instruction content before acting, which can disclose private prompts, credentials, project names, or sensitive tasks to bystanders or recordings. Because this happens before any task execution, there is no opportunity to classify or redact sensitive content first.
The skill sets Chinese as the default broadcast language in persistent guidance without requesting user consent or presenting alternatives. This can override user preferences and create confusing or privacy-impacting behavior, especially in shared environments where spoken output may be understood by unintended listeners.
The configuration examples repeatedly normalize Chinese as the default language, reinforcing a behavior change without documenting consent or preference selection. Even in examples, this can lead integrators to deploy a privacy-affecting default broadly and unintentionally.
Package name closely resembles a popular package, suggesting possible typosquatting. Attackers publish malicious packages with similar names to trick developers into installing them.
Package name closely resembles a popular package, suggesting possible typosquatting. Attackers publish malicious packages with similar names to trick developers into installing them.
A rule requiring speech output on every interaction creates a broad natural-language exfiltration channel because user prompts, agent actions, and reply summaries may be spoken aloud to anyone nearby. The requirement is especially risky because it is mandatory and global rather than context-aware.
The skill explicitly instructs modifying multiple persistent workspace files that govern future agent behavior, but it does not clearly warn that these changes will survive the current task and affect later sessions. This creates a persistence mechanism that can silently alter trust boundaries and operating behavior over time.
The examples normalize repeating user requests and response summaries in speech, which encourages developers to leak operational details and user data through audio output. Example code and docs are powerful defaults, so this pattern is likely to be copied into real deployments.
The library usage examples embed user message content and delivered results directly in spoken announcements, creating an easy copy-paste path for data leakage. Although presented as convenience examples, they operationalize insecure handling of prompt and response data.
The configuration hard-codes default_lang to zh, which makes Chinese the default output language for announcements unless changed later. This is a natural-language locale policy concern because the skill selects a specific language by default rather than offering a neutral default or explicit user opt-in at first use.
subprocess module calls execute external commands. Without careful input validation, this enables command injection.
]
try:
result = subprocess.run(
cmd,
capture_output=True,
text=True,
subprocess module calls execute external commands. Without careful input validation, this enables command injection.
return False
try:
result = subprocess.run(
player_cmd + [audio_file],
capture_output=True,
timeout=60
subprocess module calls execute external commands. Without careful input validation, this enables command injection.
# 3. edge-tts 可用性(快速检查)
try:
result = subprocess.run(
[sys.executable, "-m", "edge_tts", "--help"],
capture_output=True,
timeout=5,
This code file contains user-facing natural-language strings and module documentation that default to Chinese for key descriptions and later command feedback, without offering the user a choice of language or locale. The policy specifically calls for flagging language/locale constraints when a skill forces a specific language without opt-in.
The docstring states mandatory announcement rules entirely in Chinese and points to a rule set the script says must be followed every interaction. In addition, the main API defaults to lang="zh", which creates a language/locale default that may override user preference rather than offering a neutral choice or explicit opt-in.
Skill enumerates or reads other installed skills. Access to other skills' SKILL.md files or the skills directory reveals prompt instructions, capabilities, and secrets that should be invisible to peer skills.
complete: "上传完成"
error: "网络连接失败"
详细规则见: skills/beep-skills/SKILL.md
"""
import os
subprocess module calls execute external commands. Without careful input validation, this enables command injection.
logger.debug(f"尝试生成 ({attempt}/{MAX_RETRIES})...")
result = subprocess.run(
cmd,
capture_output=True,
text=True,
subprocess module calls execute external commands. Without careful input validation, this enables command injection.
return False
try:
result = subprocess.run(
player_cmd + [audio_file],
capture_output=True,
timeout=60,
The script sets the announcement language to "zh" by default when the user does not provide a language argument. This enforces a specific locale choice rather than prompting for or preserving user preference, which is a natural-language policy concern under the language/locale rule.
The top-level docstring mixes English with Chinese ("小喇叭") in a user-facing description. Under the stated policy, forcing or assuming a specific language without opt-in can be a natural-language policy concern when no choice or justification is provided.
The module documentation says it falls back to a 'shell script', which suggests a script-based fallback mechanism. In the code, fallback behavior is implemented via direct subprocess execution of python -m edge_tts and platform player binaries like afplay, mpg123, ffplay, or aplay, with no shell script involved.
Dynamic getattr() with a non-literal attribute name can access arbitrary object attributes, potentially bypassing access controls.
# 全局配置
_config = AnnouncementConfig.load()
setup_logging(getattr(logging, _config.log_level))
# ============ 统计与监控 ============
class StatsCollector:
Dynamic getattr() with a non-literal attribute name can access arbitrary object attributes, potentially bypassing access controls.
# 全局配置
_config = AnnouncementConfig.load()
setup_logging(getattr(logging, _config.log_level))
# ============ 统计与监控 ============
class StatsCollector:
The top-level docstring describes this as a 'PyGame Version' using pygame.mixer for playback, which suggests a narrowly scoped local playback implementation. In reality, the script also synthesizes speech by spawning python -m edge_tts and falls back to OS-level media players, so the documentation understates and partially misrepresents the implemented behavior.
Dependency has known vulnerabilities (CVEs). Using packages with unpatched security flaws exposes the environment to known exploits.