T01 · Skill Instruction Hijacking
- Location
SKILL.md:104- Finding
Untrusted Feed Titles Can Indirectly Hijack LLM Instructions
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
This skill coherently generates an AI-news digest from Feedship data, but users should understand its recurring schedule option and report-integrity risks from untrusted feed content.
Install only if you want a Chinese AI-news digest tied to a local Feedship setup. Review the cron schedule before enabling it, and prefer running it in an isolated account or session. Treat generated links and summaries as untrusted feed-derived content until the package escapes Markdown, validates URL schemes, and uses a private temporary directory instead of fixed /tmp filenames.
SKILL.md:104Untrusted Feed Titles Can Indirectly Hijack LLM Instructions
scripts/replace_refs.py:41Feed-Controlled Values Are Inserted into Markdown Without Escaping or URL Validation
SKILL.md:72Predictable Shared Temporary Files Permit Local Symlink and Data-Tampering Attacks
Referenced artifact was not completely inspected
从 `references/prompt.md` 读取提示词模板,拼接过滤后的标题列表:
Referenced artifact was not completely inspected
从 `references/prompt.md` 读取提示词模板,拼接过滤后的标题列表:
Skill contains instructions that could directly expose system prompts, internal rules, or hidden instructions to users or external parties.
---
## Output Rules
- **语气**:有远见、外科手术般精确、高度务实。无废话。
- **格式**:使用 **Markdown**,每个主要分析模块以 `## 具体主题 | 事件代号` 作为标题,**标题中不得出现 `${编号}`**。后接长篇段落或要点列表。**粗体**强调关键概念。
- **正文引用强制要求**:每个分析段落的正文叙述中(不是在列表项中),至少包含 1-2 处 `${编号}` 引用来支撑论点。**禁止把所有引用只放在列表项中,正文段落零引用。**
Skill contains instructions that could directly expose system prompts, internal rules, or hidden instructions to users or external parties.
---
## Output Rules
- **语气**:有远见、外科手术般精确、高度务实。无废话。
- **格式**:使用 **Markdown**,每个主要分析模块以 `## 具体主题 | 事件代号` 作为标题,**标题中不得出现 `${编号}`**。后接长篇段落或要点列表。**粗体**强调关键概念。
- **正文引用强制要求**:每个分析段落的正文叙述中(不是在列表项中),至少包含 1-2 处 `${编号}` 引用来支撑论点。**禁止把所有引用只放在列表项中,正文段落零引用。**
The skill invokes environment-dependent commands and relies on variables such as TODAY_ARTICLES and shell-expanded paths, but it does not declare any tool scope or execution permissions. This creates an under-specified trust boundary: an agent may execute shell/Python behavior and access environment data without explicit authorization controls, increasing the chance of unintended capability use.
The description includes broad trigger phrases such as generic daily-briefing or summary requests, which may cause the skill to activate outside the user's intended scope. Unintended activation is dangerous here because activation leads to command execution, file access, and LLM-driven processing rather than a simple text-only response.
The manual trigger list contains ambiguous phrases like 'daily digest' and '生成简报' without constraints, making accidental or adversarial invocation more likely. Because this skill performs local command execution and reads local prompt files, even mistaken activation expands the attack surface and can trigger unnecessary processing.
The workflow asks the agent to run inline shell and Python snippets, create and overwrite files under /tmp, and compose prompts from local files, which is broader than necessary for a digest-generation skill. If article content, local files, or execution context are manipulated, this wider execution surface can be abused for prompt injection, data exposure, or unsafe command execution chains.
The prompt specifies Output Language: ZH, which imposes a specific language/locale requirement. Although templated, the current file explicitly sets Chinese as the output language and does not state that the user may choose another language or opt in to this constraint.
The script’s docstring, usage guidance, and operational notes are entirely in Chinese, and its warning/output strings are likewise Chinese-only. Under the policy, forcing a specific language without user opt-in is a natural-language locale violation unless the constraint is explicitly justified, which is not present here.