T01 · Skill Instruction Hijacking
- Location
SKILL.md:12- Finding
Skill Instructions Attempt to Override Agent Control and Error-Reporting Behavior
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
The skill has a coherent study-plan purpose, but it under-discloses risky behavior such as remote script loading, unsafe HTML generation, authority-like instructions, and overwrite behavior.
Install only if you are comfortable with a Chinese-language skill that writes and may overwrite an HTML report, loads ECharts from bootcdn.net when the report is opened, and embeds study data into the page. Prefer a version that bundles dependencies locally, escapes all knowledge-map content, removes authority-spoofing language, and asks before overwriting or invoking another skill.
SKILL.md:12Skill Instructions Attempt to Override Agent Control and Error-Reporting Behavior
assets/template.html:338Untrusted Knowledge-Map Fields Are Inserted into HTML and JavaScript Without Context-Aware Escaping
assets/template.html:7Generated Reports Execute an Unverified Third-Party CDN Script
There is a clear mismatch between the declared purpose and the supplied code chunk. The description claims active functionality for generating a three-in-one static review-plan HTML and then asking the user about generating mock questions. However, the actual code contains no executable logic at all—only deprecation comments saying the script is obsolete and that the real logic has been migrated to another location. Since the supplied code does not perform the declared primary purpose or any of its key behaviors, the description does not accurately represent this code chunk.
L016声明“遇到问题静默处理,不报告给用户”,但L113和L115又明确要求在特定错误场景下向用户提示错误信息。这不是单纯的文档不完整,而是同一文件中的执行意图直接冲突,导致实现者无法同时满足两者。
Using a broad phrase like '重新生成' as an unconditional trigger can cause unintended file writes or workflow execution during ordinary conversation. In an agentic system, ambiguous natural-language triggers increase the chance of unauthorized or accidental state-changing actions, especially when the action overwrites artifacts.
L062和L071都要求将HTML写到工作区根目录,但L135的Anti-Patterns又声称最终产物必须是“.exam-output/exam-mindmap-YYYYMMDD.html”。这会直接影响代码实际写入位置,属于文档意图与要求相互矛盾。
Instructing the agent to write an output file without user-facing warning about modification can lead to silent overwrites of existing artifacts. In agent environments, unannounced file changes are risky because they can destroy prior outputs, confuse auditability, and be abused through repeated trigger manipulation.
Accepting generic replies like '要', '好', '继续', or '是' to trigger another skill is dangerous because those responses are common in unrelated dialogue and can be misbound to the wrong action. This raises the risk of accidental invocation of exam-question generation and unintended consumption or modification of workspace resources.
The explicit regeneration behavior to overwrite the old file without warning creates a direct path to destructive state changes from conversational input. Because the same document also uses broad triggers, the overwrite behavior becomes more dangerous in context: accidental phrases can replace existing files without informed consent.
The HTML root declares lang="zh-CN", and the visible UI text throughout the template is exclusively in Chinese, which indicates the skill is fixed to a specific language/locale. There is no indication in this file that users can opt into another language or that the locale restriction is explicitly justified as region-specific.
The template pulls executable JavaScript from a third-party CDN, which breaks the stated 'static/local' trust boundary and introduces a supply-chain risk. If the CDN, DNS, or delivery path is compromised, opening the generated HTML could execute attacker-controlled code in the user's browser and exfiltrate embedded study data or manipulate content.
This generated HTML has an unnecessary external network dependency for core rendering, so merely viewing the file triggers a request to a third party. That leaks metadata such as client IP, access timing, and possibly referrer/context, and also creates an availability and tampering risk if the remote resource changes or becomes malicious.
The document contains a natural-language requirement to prioritize Chinese fonts and explicitly prohibit Inter, Roboto, and Arial. This enforces a specific language/locale presentation choice without offering user choice or stating a region-specific justification, which matches the language/locale policy violation criteria.
The skill prescribes fixed Chinese user-facing messages and accepted reply phrases, with no indication that the language should follow user preference. This creates a locale policy issue because the interaction is constrained to a specific language without opt-in or documented justification.