T01 · Skill Instruction Hijacking
- Location
workflows/reg.md:65- Finding
Persistent Agent-Wide Instruction Hijacking Through AGENTS.md Modification
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
This skill is a real Feishu task automation workflow, but it also installs persistent automation, changes workspace-wide agent routing, and can publish local file-derived summaries to Feishu without enough user control.
Install only if you intentionally want this skill to create and update Feishu tasks, run recurring OpenClaw jobs, update scheduler messages, and change workspace routing behavior. Before use, review the cron jobs, AGENTS.md changes, hidden scheduled-message block, and daily.json upload path; use a constrained workspace that contains only intended journal files and avoid sensitive date-named Markdown files under the app root.
workflows/reg.md:65Persistent Agent-Wide Instruction Hijacking Through AGENTS.md Modification
workflows/reg.md:31Installation of Recurring Cross-Session Background Jobs
references/task-decision-rules.md:127Concealed Control Instructions Embedded in Scheduled Messages
workflows/polled-task-execution.md:75Shell Command Injection Risk During Cron Message Updates
scripts/build_daily_json.py:195Overbroad Local File Discovery Followed by Remote Profile Publication
This is the same core issue in Chinese: the skill claims to be the main entry for task creation, polling, registration, and daily.json generation, but the observed capabilities are much narrower. Overstated functionality in a routing skill can mislead reviewers and users, causing unsafe reliance, over-permissioning, and hidden access to local application context.
This is the same core issue in Chinese: the skill claims to be the main entry for task creation, polling, registration, and daily.json generation, but the observed capabilities are much narrower. Overstated functionality in a routing skill can mislead reviewers and users, causing unsafe reliance, over-permissioning, and hidden access to local application context.
The skill declares no explicit tool scope even though it is documented as using file read/write and environment-backed capabilities. In an agent ecosystem, missing permission boundaries can let the skill access broader resources than callers expect, increasing the blast radius if downstream workflow content is compromised or misrouted.
The skill enables implicit invocation without any visible trigger constraints, exclusions, or narrowing conditions in this file. Because this agent can create tasks, process unfinished tasks, perform registration/initialization flows, and generate or refresh daily.json, broad auto-routing increases the chance of unintended activation and execution on ambiguous user input, which could lead to unauthorized task creation, unwanted workflow actions, or disclosure/modification of operational summaries.
The entire rule file is written as mandatory operational guidance in Chinese, including required user-facing reply behavior, but it does not indicate that language selection depends on user preference or locale. Because these instructions govern final user replies and confirmations, they effectively impose a specific language without opt-in.
The example explicitly instructs creation of a scheduled task whose trigger message contains hidden, non-user-visible processing instructions and references automatic execution via a polling workflow. That creates a transparency and consent problem: background jobs may later act on Feishu messages, tasks, or calendar context without the user being clearly informed at setup time, increasing the risk of confused-deputy behavior or overbroad data access.
The workflow explicitly sends the generated daily.json content to an external Feishu task agent via update_profile, but the skill text provides no privacy notice, consent gate, or data-classification limits before exporting potentially sensitive project summaries. Because daily.json is derived from the latest日报 and may include internal plans, deliverables, and upcoming work, this creates a real data-sharing risk rather than a purely documentation issue.
The entire workflow content is authored in Chinese and presents the procedure as the required operating language, with no opt-in, fallback, or note that the skill is limited to a Chinese-speaking environment. Under the policy, language constraints should either be optional for the user or explicitly justified as region- or compliance-specific.
The workflow goes beyond Feishu task execution and directs the agent to modify external scheduler state by discovering cronjob IDs, rewriting payload messages, and issuing openclaw cron update. This expands the skill's authority from task processing into host automation control, creating a cross-boundary side effect that could alter future execution behavior or retarget scheduled jobs if task data or matching logic is wrong or manipulated.
Using host CLI commands to enumerate all cronjobs and update one is broader than the stated purpose of handling Feishu tasks, and it grants unnecessary access to scheduler configuration. Enumeration plus free-form message replacement increases the chance of touching unrelated jobs, leaking scheduler metadata, or enabling abuse if an attacker can influence the selected cronjob or replacement content.
The workflow performs scheduler configuration changes through shell commands without any user-facing warning, approval checkpoint, or audit-oriented disclosure. Silent modification of cron configuration is risky because it changes persistent automation behavior outside the visible Feishu task domain, making misuse or mistakes hard for users to detect until later executions occur.
The workflow creates persistent cron jobs and modifies AGENTS.md without any explicit user-facing warning or confirmation that it will change the workspace and schedule ongoing automated actions. This can surprise users, create durable side effects, and establish recurring execution paths that continue after the initial request, increasing the risk of unauthorized persistence.
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 * * * *` |
Hard-coding the timezone to Asia/Shanghai without user opt-in can cause scheduled jobs to run at unintended local times, which may lead to missed deadlines, unexpected automation, or execution outside approved operating windows. In a task automation skill, scheduling semantics are security-relevant because they control when persistent automated actions occur.
The registration workflow changes the workspace-level AGENTS.md to enforce global routing behavior for future user inputs, which is far broader than simple Feishu task registration. This creates persistent behavior modification and can silently redirect unrelated future requests into this skill, expanding privilege and control beyond the user’s immediate registration intent.
The display and description text include Chinese while the default prompt is in English, but the manifest does not say whether language is selected based on user preference. This can create a locale-policy issue if the skill presents or enforces a language without explicit opt-in or documented justification.
The document is entirely written in Chinese and does not state that the language is optional, user-selectable, or limited to a China-specific workflow. Under the policy, forcing a specific language without user opt-in can be a natural-language policy violation.
The natural-language content throughout the file is Chinese-only, including headings, examples, and expected behavior, with no indication that users may choose another language or that the skill is intentionally limited to a Chinese-speaking environment. Per the policy, forcing a specific language without opt-in is a natural-language policy concern.
This Python file contains multiple inline comments written only in Chinese while the surrounding code and CLI interface are otherwise English-oriented. That creates a language/locale constraint in the natural-language content of the skill without offering any choice or documenting a justified region-specific requirement.
All natural-language instructions in this workflow are Chinese-only, which can amount to a language policy violation when no user opt-in or alternative locale is offered. The file does not document that the workflow is intentionally region-specific or provide any language choice.
The overview and dependencies state that the full registration/initialization flow relies on OpenClaw built-in Tool capabilities for subsequent registration operations. But the actual steps culminate in directly calling the feishu_task_agent tool for register, which is a distinct tool dependency not reflected by the earlier wording and creates a documentation-to-execution contradiction.