Security audit
Security checks for vulnerabilities and agentic risk
This skill is a documentation-only teaching aid that tells an agent to explain agent architecture and, when asked about a specific implementation, read relevant local source code before answering.
Install this if you want your agent to answer architecture questions by inspecting relevant source files. Be aware that specific implementation questions may cause the agent to read local agent source code and quote file paths or line references, so avoid using it for private codebases if that disclosure would be inappropriate.
Most of the README content is presented in Chinese, which effectively imposes a specific language on users reading the skill documentation. The file does not offer an alternative language, user choice, or note that the skill is intentionally region- or language-specific.
The README defines activation using natural-language conditions such as 'general knowledge and basic concepts' and 'current Agent's specific architecture implementation,' but it does not provide precise criteria for borderline cases. Without explicit constraints or negative examples, the skill may inconsistently decide whether to activate and whether source-reading is required.
The README says users can 'directly ask your Agent architecture questions' and that the skill 'will automatically activate,' but it does not define a concrete trigger list, exclusions, or negative examples. This ambiguity is reinforced by broad conditions like 'when asked about architecture questions' elsewhere, which could overlap with many normal conversations about agents and cause unintended invocation.
The skill activates on broad prompts about 'Agent architecture' or 'how a specific Agent works', which can overlap with ordinary product, codebase, or system-design questions. In an agent-routing context, this can cause the skill to take over conversations unnecessarily and push the model into self-referential code-reading behavior, increasing the chance of scope hijacking or unintended disclosure of internal implementation details.
The example trigger phrase 'What's the architecture of this system?' is highly ambiguous and may match many normal engineering discussions unrelated to agent internals. Because the skill's Layer 2 behavior explicitly encourages searching the agent's own source code, an overly generic trigger can route unrelated queries into unnecessary code inspection and explanation of internal structure.
The trigger list includes very generic terms such as "tools," "memory," "config," and "provider," which are common across many unrelated user requests. In a skill-routing system, this can cause unintended activation, leading the agent to apply the wrong behavior or disclose architecture-focused responses when the user did not ask for them.
The architecture map states that tool execution happens in a 'sandboxed environment,' which is a concrete safety/property claim rather than a neutral pointer to source layout. In this file, no supporting implementation detail or qualification is provided, so the documentation makes a stronger assertion about behavior than the referenced material substantiates.
The file includes Chinese-language trigger terms alongside English ones, which implies language-specific behavior, but it does not explain whether multilingual activation is optional, user-selected, or universally enforced. This can create an implicit language/locale policy ambiguity under the language-choice rule.
No suspicious patterns detected.
