T01 · Skill Instruction Hijacking
- Location
workflows/reg.md:66- Finding
Persistent Modification of Global Agent Routing Instructions
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
The skill is a real Feishu task automation skill, but it makes lasting changes that can steer future agent behavior and run scheduled workflows without enough visible user control.
Review this skill carefully before installing. It should only be used if you intentionally want Feishu task automation that can create and update tasks, run recurring OpenClaw cron jobs, and modify AGENTS.md routing behavior. Ask the publisher to add explicit permission declarations, separate confirmation for each persistent change, visible scheduler metadata instead of hidden prompt text, and clear uninstall or rollback instructions.
workflows/reg.md:66Persistent Modification of Global Agent Routing Instructions
workflows/reg.md:30Registration Installs Recurring Cross-Session Scheduled Jobs
references/task-decision-rules.md:127Hidden Agent-Control Instructions Embedded in Recurring Task Messages
代码行为非常明确且单一:遍历应用目录查找形如 YYYY-MM-DD.md 的日报文件,选择最新候选,解析 Markdown 内容,基于关键词规则抽取近期亮点和明日关注,最后写入 app_root/daily.json。它访问的是本地文件系统,没有任何飞书 API、任务系统、网络请求、注册流程、轮询调度或任务执行相关代码。因此,虽然声明中的第(4)项与代码一致,但声明把该技能描述为统一处理四类请求的总入口,而当前代码块只覆盖其中一小部分。按评估标准,这属于声明目的与实际行为存在实质性差异:主要能力被过度声明,且实际主用途只是 daily.json 生成脚本。
The description presents a broad 'total entry' skill for multiple user-facing workflows, but the code chunk only implements a specific backend helper for resolving creator/member payloads for Feishu tasks. It validates IDs, determines assignee/follower roles, merges runtime context, optionally reads OpenClaw config to find an app_id, and prints normalized JSON. Those behaviors may support task creation internally, but they do not themselves implement the declared primary functions such as task orchestration, polling unfinished tasks, agent registration, or daily.json extraction. This is therefore a material description-to-behavior mismatch.
The code does align with one subset of the description: handling inputs from polling of unfinished tasks and deciding whether they should proceed. However, the declared purpose presents this skill as a unified entry point for four major categories of behavior, including task creation/orchestration, agent registration/initialization, and daily.json generation from reports. None of those behaviors appear in the supplied code chunk. The actual code only performs decision logic for whether a polled task should execute, with JSON parsing and timestamp validation. This is a materially narrower and different implementation than the declared multi-function gateway, so the description does not accurately represent what this code chunk actually does.
Without declared permissions the skill's intent is opaque and cannot be validated.
The default prompt routes a wide range of intents through a single skill and the policy allows implicit invocation, but it does not define sufficiently narrow activation boundaries. This can cause the skill to trigger on ambiguous user input and perform task orchestration, registration, or daily.json generation when the user did not clearly request those actions, increasing the risk of unintended side effects or overbroad tool use.
All user-facing instructions in this skill file are presented only in Chinese, and there is no indication that the skill is region-specific or that users may choose another language. Under the stated policy, forcing a specific language without opt-in is a natural-language policy violation.
The document is written as normative skill behavior entirely in Chinese, establishing operational rules and user-facing reply requirements in that locale. There is no indication that the skill offers a language choice, user opt-in, or that the Chinese-only constraint is justified as a region-specific requirement.
The scheduled-task rules instruct the agent to embed hidden control text into the cron message and to generate future content using broad user context such as Feishu messages, tasks, and calendar data. Because that payload is explicitly marked as not shown to the user and later re-consumed as natural-language instructions, it creates a covert prompt channel that can trigger overbroad data access or disclosure without transparent user consent or tight scoping.
The examples direct the agent to create OpenClaw cron jobs, which introduces a scheduling/execution capability beyond the four declared request types in the skill metadata. That mismatch is dangerous because it can cause the agent to perform persistent automated actions that users and reviewers may not expect, weakening least-privilege boundaries and enabling unintended recurring task execution.
The example requires inserting hidden, non-user-displayed directives into the scheduled message, including internal processing instructions and task identifiers. Concealed control metadata is risky because it can steer later agent behavior without user awareness, creating a channel for covert instruction passing, confused-deputy behavior, and reduced auditability of automated actions.
The workflow instructions are entirely written in Chinese and line L17 requires the agent to '明确告知' the result, implying a fixed language/locale behavior. There is no indication that the user may choose another language or that the Chinese-only constraint is optional or region-specific.
The file title and instructions are written entirely in Chinese and define the workflow behavior without offering any user language choice or indicating that the workflow is limited to a Chinese-language environment. Under the policy, forcing a specific language without opt-in is a natural-language policy violation.
The workflow expands from Feishu task processing into shell-level cron administration by listing cronjobs and updating their payload messages based on task data. This creates a cross-boundary capability where task content can influence scheduler state, raising risks of unauthorized persistence changes, misrouting future executions, or tampering with other scheduled jobs if identifiers or message mapping are wrong or manipulated.
The workflow instructs the agent to execute shell commands for cron list/update even though the skill's stated purpose is Feishu task handling. Giving an LLM-driven workflow operational control over scheduler configuration increases attack surface substantially, especially because it derives the updated message from runtime data and performs command-level changes that can affect future autonomous behavior.
The document claims the registration flow only triggers on explicit user requests, but the AGENTS.md content it installs later forces automatic pre-routing of many future requests through feishu-task-agent. This mismatch is dangerous because it conceals a persistent change in control flow: users and operators may believe the skill is narrowly triggered when it actually establishes broad automatic interception after registration.
The workflow instructs creation of cron jobs and modification of AGENTS.md without an explicit warning that persistent system-level changes will be made. Silent persistence is risky because it can alter future agent behavior, schedule autonomous execution, and continue operating after the initiating conversation, all without clear user awareness or approval.
Large whitespace padding was detected (a block of blank lines or a long run of spaces). This can push injected instructions below or to the right of the visible area so a human reviewer never sees them while the agent still reads them. Manual review of the hidden content is recommended.
- 判断规则固定为:返回结果中 `name` 等于对应 `Task ID`,即视为该任务已存在,必须跳过,不得重复创建。
- 根据以下明确任务列表,按顺序逐个执行“先查是否存在,再决定创建/跳过”:
| 任务唯一 ID (Task ID) | 任务描述 (Description) | Cron 表达式 |
| --------------------------------- | -------------------------------------------------------------------------------------------------- | ---------------------- |
| feishu-task-daily-report-genarate | 使用技能 `feishu-task-agent`,按 `workflows/agent-profile.md` 的规则在当前应用目录生成或更新 `daily.json`,输出“Todo_Agent_RecentHighlights_Field”和“Todo_Agent_TomorrowsFocus_Field” | `30 23 * * *` |
| get-agent-unfinished-tasks | 使用技能 `feishu-task-agent`,流转到 `workflows/polled-task-execution.md` 执行未开始任务轮询流程 | `*/10 * * * *` |
The workflow hard-codes the Asia/Shanghai timezone for scheduled tasks without checking user locale, workspace settings, or consent. This can cause tasks to run at unintended times, leading to missed deadlines, unexpected automated activity, or disclosure/processing of daily reports outside the user's expected operating hours.
The workflow directs modification of the host agent's AGENTS.md to install broad routing rules that affect all future user inputs, not just Feishu-task registration. This creates persistent behavioral changes and privilege expansion beyond the user's likely expectation for a one-time registration flow, enabling the skill to hijack or bias later interactions toward its own tooling.
This is a markdown file, so SQP-2 applies to omissions in the skill description. Line L03 states that the script always writes a JSON file at the application root, but the document does not include any warning that the skill modifies user/project files or may overwrite existing daily.json content.
This markdown file states that changes will overwrite daily.json and may create an initialization placeholder file if none exists. Because this is a markdown skill document and it describes file-affecting behavior without any explicit warning about modifying user data on disk, it matches the missing user warning criterion.
Natural-language policy requires avoiding forced language/locale without user opt-in or documented justification. This file presents all instructions and examples only in Chinese, with no indication that the skill is intentionally region-specific or that users may choose another language.
The documentation explicitly permits passing a user identifier (sender_open_id) via environment variables but provides no warning about exposure risks such as process inspection, shell history, inherited environments, CI logs, or accidental debugging output. While open_id is not a secret credential, it is still user-linked identifier data, and normalizing env-var transport without handling guidance can lead to unnecessary privacy leakage and weak operational hygiene.
The code contains multiple natural-language comments in Chinese only, such as the task-handling explanations around L094-L135. This imposes a specific language for maintainers or reviewers without any indication of language choice or documented locale constraint, which matches the language/locale policy concern.