T01 · Skill Instruction Hijacking
- Location
SKILL.md:104- Finding
Remote Team Instructions Can Hijack Agent Goals and Communication
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
The skill is a plausible team-collaboration tool, but it gives remote team messages and Git/network operations too much under-scoped authority.
Install only if you trust the team service and repository workflow. Use an approved HTTPS endpoint, avoid sending secrets in team messages, do not embed Git tokens in URLs, keep config and task files in a dedicated workspace path, and require explicit user approval before accepting remote tasks or pushing changes.
SKILL.md:104Remote Team Instructions Can Hijack Agent Goals and Communication
scripts/send_message.js:55Arbitrary Plaintext HTTP Endpoint Enables Data Disclosure and Server-Side Request Abuse
SKILL.md:130Git Access Token Is Placed in the Clone URL
scripts/join_team.js:43Caller-Controlled Configuration Path Permits Arbitrary File Overwrite
声明描述的是一个较完整的“团队协作技能”,涵盖初始化注册、消息通信和状态同步三类核心能力。但提供的代码块只完成加入团队这一初始化步骤:校验参数、发起 HTTP POST 请求加入团队、并把配置写入文件。代码没有任何消息发送/接收逻辑,也没有轮询、订阅、同步状态或维护团队共享状态的实现。因此,代码行为仅覆盖了声明中的一部分,且缺少两个被明确列为“主要功能”的能力,构成描述与实际行为不一致。
The skill directs agents to create, clear, and continuously update a local tracking file without warning that this modifies the workspace. Unannounced file writes are risky because they can overwrite user data, create persistence artifacts, or leak intermediate reasoning/plans into project files unintentionally.
The documentation broadens the skill from team communication into Git-based remote file sharing and repository workflows, which introduces network access, remote writes, and possible propagation of sensitive workspace content. In this context, an agent following the skill could modify or publish data to external infrastructure not strictly necessary for the declared function.
The skill prescribes clone/push workflows and authenticated remote access without any warning about network transmission, remote side effects, or the possibility of publishing sensitive local changes. In a multi-agent setting this is more dangerous because agents are instructed to push before communicating, normalizing remote writes as a prerequisite and increasing the chance of unintended disclosure or repository tampering.
The skill instructs agents to use Git credentials from environment variables and embed them into clone URLs, expanding the trust boundary beyond team messaging into authenticated remote access. This can cause accidental credential exposure through command history, logs, process listings, or unintended use of secrets for unrelated repositories/services.
The entire skill guidance is written as a Chinese-only operational instruction, with no indication that the user may choose another language or that the skill is limited to a Chinese-specific context. Under the policy, a fixed language or locale without opt-in can be a natural-language policy violation.
The guideline instructs the lead agent to clear task_plan.md before starting a new task, which can delete prior task history, evidence, or user-relevant work products without warning or confirmation. In a team-coordination skill, this is risky because task files are part of operational state and their silent removal can cause loss of auditability, progress tracking, and recovery data.
The file is written entirely in Chinese and presents mandatory behavioral instructions without offering any language choice or opt-in. Under the policy, forcing a specific language or locale without user choice is a natural-language policy violation unless clearly justified as region-specific, which is not stated here.
This code includes natural-language comments, errors, and console output exclusively in Chinese, such as parameter errors and send-status messages. That can violate a language/locale policy when the skill does not offer user opt-in or an alternative language.
This file contains multiple user-facing comments and console messages in Chinese, including required-parameter errors and success/failure notices. Because the script does not offer any language selection or document a justified locale restriction, it violates the language/locale policy criterion.