T01 · Skill Instruction Hijacking
- Location
prompts/executor-main.md:26- Finding
Untrusted User, Web, and Memory Content Is Interpolated into Subagent Prompts
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
This is a disclosed music-generation skill that uses kie.ai, web search, subagents, and local preference memory in ways that match its stated purpose, though users should understand the privacy and persistence tradeoffs.
Install only if you are comfortable sending prompts, lyrics, style tags, and titles to kie.ai/Suno-related services and storing liked-generation history locally. Avoid submitting secrets or sensitive personal material, leave CALLBACK_URL empty unless you control it, and periodically delete the skill's memory files if you do not want preference history reused.
prompts/executor-main.md:26Untrusted User, Web, and Memory Content Is Interpolated into Subagent Prompts
README.md:87Mutable Global Installation Instructions Create a Supply-Chain Exposure
Credentials or environment variables flow to a network sink. This is a high-confidence indicator of credential exfiltration.
print("用法: python check_task.py <task_id>", file=sys.stderr)
sys.exit(1)
resp = requests.get(
f"https://api.kie.ai/api/v1/generate/record-info?taskId={task_id}",
headers={"Authorization": f"Bearer {API_KEY}"},
timeout=20,
Credentials or environment variables flow to a network sink. This is a high-confidence indicator of credential exfiltration.
end_time = time.time() + timeout
while time.time() < end_time:
resp = requests.get(
f"{BASE_URL}/api/v1/generate/record-info?taskId={task_id}",
headers=HEADERS, timeout=20, verify=True
)
Credentials or environment variables flow to a network sink. This is a high-confidence indicator of credential exfiltration.
if CALLBACK_URL:
payload["callBackUrl"] = CALLBACK_URL
resp = requests.post(
f"{BASE_URL}/api/v1/generate",
headers=HEADERS,
json=payload,
Credentials or environment variables flow to a network sink. This is a high-confidence indicator of credential exfiltration.
raise Exception("错误: 请设置 KIEAI_API_KEY 环境变量")
url = f"{BASE_URL}/api/v1/generate/record-info?taskId={task_id}"
resp = requests.get(url, headers=HEADERS, timeout=20)
data = resp.json()
if data.get("code") != 200:
return None
Referenced artifact was not completely inspected
当运行本技能时,主代理按照 `prompts/executor-main.md` 中的指引执行:
The design sends prompts, lyrics, style tags, and related song-generation data to the external kie.ai API, but the document does not indicate any disclosure that user content leaves the system boundary. This can expose user-provided creative material or personal data to a third party without informed consent, creating privacy, compliance, and trust issues.
The architecture explicitly states that complete interaction snapshots are stored in history.json and summarized into patterns.log for later reuse, but it does not mention user notice, consent, retention limits, or safeguards. In a creative assistant, prompts and generated lyrics can contain personal, sensitive, or proprietary information, so persistent storage and reuse create a meaningful privacy and data-governance risk.
The architecture not only retains prior interactions, it also re-injects derived history into future agent rounds via HISTORY_SIGNAL. This creates a leakage path where prior user content, preferences, or sensitive details can influence later outputs unexpectedly, potentially surfacing retained information across sessions or contexts.
The README explicitly describes persistent storage of interaction history and long-term preference patterns, but does not disclose privacy, retention, or consent implications. In a creative workflow, users may submit sensitive prompts, personal preferences, or copyrighted text, so undocumented persistence increases the risk of unintended data retention and user surprise.
The README instructs users to configure a third-party API key and use the skill for music generation, but does not warn that prompts, lyrics, and related content will be transmitted to kie.ai/Suno-related external services. This creates a transparency and privacy risk because users may assume processing is local when their creative content is actually sent off-platform.
整份技能说明、状态输出和交互示例均固定为中文,且未说明是否支持其他语言或允许用户选择输出语言。根据语言/区域政策,若技能默认强制特定语言而没有用户选择或明确限定场景,属于自然语言策略风险。
The trigger phrases are broad natural-language requests such as '制作一首...' or '创作一首歌...', which can cause the skill to activate on ordinary conversation not explicitly intended to invoke this workflow. In this skill, accidental activation is more concerning because it can lead to downstream API usage, external content generation, and persistence of user inputs into memory files without a clearly deliberate opt-in.
The architecture defines complete interaction snapshots in hot storage and unlimited append-only long-term logging of preference signals, encouraging indefinite accumulation of user behavior data. This materially raises the privacy and security risk because a larger, long-lived corpus increases the blast radius of any exposure and makes secondary use of personal data more likely.
The skill explicitly instructs collecting detailed user ratings and free-form comments, then persisting them to memory to shape future outputs, but it does not present a meaningful privacy notice, retention explanation, or explicit consent flow. This creates a real privacy risk because users are nudged to disclose preferences and comments without understanding that their data will be stored and reused beyond the immediate session.
The design stores user-provided feedback and comments in natural language and reuses them in later prompt construction, which extends data use beyond the immediate task. This is dangerous because free-form feedback can contain sensitive personal information, and persistent reuse increases the chance of unintended disclosure, over-retention, or prompt-context leakage.
The example schema persists raw user ideas, ratings, and free-form comments directly in history files, which can capture sensitive or identifying content and make it available for later reuse or accidental exposure. Because these fields are stored in plain structured form, the design increases the likelihood of data leakage through logs, backups, local file access, or prompt reinjection.
The prompt imposes mandatory Chinese-specific rules and frames the skill primarily as a Chinese lyric-writing agent, while only secondarily mentioning English. This creates a locale/language constraint that is not presented as a user choice or explicit opt-in, which fits the language policy violation category.
This markdown prompt defines the agent role, instructions, and output field descriptions exclusively in Chinese, which effectively forces a specific language/locale for operation. The file does not provide any user choice, fallback, or explicit justification for restricting the skill to Chinese-language interaction.
The file’s title and all operative instructions are written in Chinese, and the required output schema is entirely Chinese-language, which effectively constrains the interaction/output language. There is no indication that the user can opt into another language or that this is a justified region-specific skill, so it conflicts with the policy against forcing a specific language without user opt-in.
The skill instructs the agent to send user-provided creative descriptions to web_search, which may disclose sensitive or proprietary user ideas to an external service without notice or consent. Even if the feature is functionally useful, silently forwarding raw prompts creates a privacy and data-handling risk, especially for unpublished creative work.
The document explicitly says representative artists are for reference only and must not be disclosed to sub-agents, yet the prompt builders pass the entire info_table unchanged into both sub-agent prompts. That creates an instruction/data isolation failure: downstream agents receive restricted reference data and may echo, transform, or overfit to named artists, undermining policy intent and increasing leakage/copyright-style mimicry risk.
The skill reads historical preferences from memory/patterns.log and later writes user preference/history data, but does not describe user notice, consent, retention controls, or deletion behavior. This creates a transparency and privacy vulnerability because users may unknowingly have creative preferences and interactions persisted across sessions.
The validator explicitly enforces '中文歌词' requirements, including minimum section line counts and a total length threshold, which constrains output language. The file does not offer user opt-in for Chinese output or explain that the skill is intentionally limited to a Chinese-language context.
The memory archiving step stores raw user ideas, generated prompts, audio URLs, and derived tags in persistent history files, creating a durable natural-language data retention path. If these files are later exposed, reused in prompts, or accessed by other agents/users, they can leak sensitive creative content and behavioral preferences beyond the original interaction.
Line L124 states that music descriptions are '英文推荐' (English recommended), which imposes a language preference in the skill guidance. Under the policy, language constraints should either be user-opt-in or clearly justified as region- or compliance-specific, neither of which is provided here.