T01 ยท Skill Instruction Hijacking
- Location
core/triggers.md:5- Finding
Mandatory Routing Rules Hijack the Agent's Current Response Behavior
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
This investment-research skill is coherent, but it asks for persistent agent control, broad automatic routing, and local research-state storage that users should review carefully before installing.
Install only in a dedicated investment-research workspace, not a global system prompt, unless you really want it to affect future finance-related conversations. Review any generated memory, coverage, bias, decision-log, checkpoint, and archive files before use, and avoid storing confidential client or firm information there. Treat batch refresh and automatic routing as high-impact workflows that should require explicit confirmation.
core/triggers.md:5Mandatory Routing Rules Hijack the Agent's Current Response Behavior
README.md:806Global Agent Configuration and Workspace Memory Create Persistent Skill Control
The trigger phrases include very broad everyday inputs like greetings, generic help requests, and short ambiguous text, which can hijack unrelated conversations and force the skill into menu mode. Overbroad activation increases the chance of unintended routing and, in this skill, can funnel users toward persistence or setup behaviors they did not request.
The listed phrases are broad enough to overlap with normal conversational language, so ordinary questions can activate specialized skills unintentionally. In this skill, broad triggers such as generic 'how is X' style phrasing increase the risk of workflow hijacking, where the agent abandons normal answering and enters a mandatory skill path the user did not clearly request.
This rule explicitly tells the model to trigger on semantic similarity rather than clear literals, which materially widens activation conditions and makes false positives likely. In an investment-research skill, that can silently force structured workflows, tool calls, or constrained outputs even when the user asked a casual or unrelated question, reducing user control and increasing the chance of inappropriate downstream actions.
The skill explicitly instructs the model to inspect and use user-specific memory files such as biases.md and related workspace artifacts. In an LLM environment, this can surface prior sensitive notes, internal research history, personal identifiers, or confidential firm context into new outputs beyond the user's immediate intent, increasing the risk of inadvertent data disclosure across tasks.
The instruction says users must copy the enable prompt into ~/.claude/็ณป็ปๆ็คบ่ฏ, and later repeats this Chinese-labeled path as the expected location. This imposes a specific locale-specific UI path without opt-in or an alternative for users whose tool or environment is not localized in Chinese.
Skill enumerates or reads other installed skills. Access to other skills' SKILL.md files or the skills directory reveals prompt instructions, capabilities, and secrets that should be invisible to peer skills.
### ่ทฏๅพ A ยท ๆๅฐๅฎ่ฃ
๏ผ1 ๅ้๏ผ
ๆ `skills/sm-master/SKILL.md` ๅคๅถๅฐไปปไฝ AI ๅทฅๅ
ท็็ณป็ปๆ็คบ่ฏ้ใ1 ไธชๆไปถ่งฃๅณ 80% ้ๆฑใ
### ่ทฏๅพ B ยท ๅฎๆดๅฎ่ฃ
๏ผ5 ๅ้๏ผ
The manifest says users can 'Enter a number or describe your task to begin,' which permits arbitrary natural-language descriptions as activation input. Because the file does not define boundaries or negative examples for what counts as an in-scope task, this can cause unintended invocation from ordinary investing-related conversation.
The routing rules state that if the user describes a task with no number, the skill should 'auto-route via sm-autopilot logic.' This is a vague activation condition because it does not specify how the system distinguishes intentional skill invocation from general discussion, nor does it provide exclusions or negative examples.
The file states in Chinese that this is the first file read each new session and presents the skill framework entirely in Chinese, effectively imposing a specific language/locale on the interaction flow. There is no indication that users may choose another language or that the locale restriction is intentional and justified.
Line L001 titles the acceptance checklist as '่พๅบ้ชๆถๆธ ๅ' and the entire file requires outputs in Chinese, including fixed Chinese section names and phrasing. Under the policy, forcing a specific language without user opt-in is a natural-language locale violation unless the regional constraint is explicitly documented and justified, which is not present here.
This markdown instruction applies a hard language constraint: all queries must use Chinese natural language. Under the policy, forcing a specific language without offering user choice or clearly documenting a justified region-specific limitation is a natural-language policy violation.
The checkpoint format explicitly persists full completed sections, in-progress content, and fetched-data history to workspace files. That creates durable local retention of potentially sensitive research notes and source history beyond the transient interaction implied by the public research menu, increasing exposure if the workspace is later accessed, synced, or reused across tasks.
The skill instructs the agent to create, modify, archive, and delete workspace state files such as .checkpoint and .task-pulse as part of normal operation. This is a real stateful side effect not reflected in the user-facing description, so users may unknowingly grant the skill file-management capabilities that can alter local workspace state and erase evidence of prior runs.
The documented resume flow includes automatic deletion of the checkpoint after completion, but the skill metadata shown to users does not disclose that destructive behavior. Lack of notice is dangerous because users may expect resumability or auditability, while the skill silently removes recovery state and potentially impairs troubleshooting or review.
The resume trigger accepts broad natural-language phrases such as 'ๆฅ็ไธๆฌก็', which can match ordinary conversation and unintentionally reactivate prior task state. In a stateful skill that reads checkpoint files and continues workflows, ambiguous activation can cause unintended data loading, file writes, or continuation of stale research without a deliberate user command.
The entire skill guidance is written as mandatory Chinese-language policy text, and there is no indication that users may choose another language or that the skill is limited to a Chinese-language or region-specific context. Under the policy, language constraints without user opt-in or clear justification are natural-language policy violations.
The file is entirely written in Chinese and contains no indication that users may choose another language or that the skill is intentionally restricted to Chinese-language operation. Under the policy, forcing a specific language without user opt-in is a natural-language policy violation unless the locale constraint is clearly documented and justified.
The menu content and surrounding instructions are written to be directly copied in Chinese, and the file says the LLM must directly display this menu when triggers match. There is no indication that the user can choose another language or that Chinese is a justified locale constraint.
The condition based on input being fewer than five characters and lacking keywords is ambiguous and likely to misclassify benign user messages. While not code-execution dangerous by itself, it enables unwanted takeover of the conversation flow and compounds the broader activation problem in this skill.
The menu file expands from simple routing into instructing the agent to generate and persist multiple workspace files, which materially changes user state outside the narrow role of a menu entrypoint. Because these writes are triggered from conversational flow and not framed as an explicit external side effect requiring confirmation, the skill could cause unintended persistence of sensitive investment preferences, watchlists, or working notes.
The skill instructs the agent to perform installation or setup-style actions such as running a command, even though the file is presented as a menu entrypoint. This creates an unsafe precedent where ordinary menu interaction can lead to execution-adjacent behavior or system modification beyond what a user reasonably expects from selecting a research tool.
The initialization flow directs creation of several persistent user files immediately after collecting answers, but it does not require a clear warning about filesystem modification, storage location, or what data will be written. In a skill that may capture role, coverage universe, and tracked securities, silent persistence can create privacy, integrity, and user-trust risks.
The title and the document content prescribe a Chinese-language archive naming standard, including Chinese company names and Chinese directory examples, as the required norm. This is a natural-language locale policy constraint without any opt-in or alternative for users who may operate in other languages/locales.
The title and instructions are presented as a mandatory Chinese postamble ('ๅผบๅถ็ปๆๅๆต็จ'), indicating a fixed language/locale expectation. Under the policy, forcing a specific language without user opt-in is a natural-language policy violation unless the locale restriction is explicitly justified.
The markdown directs automatic checkpoint and task-state writes during normal execution without a user-facing warning that local files will be created and updated. This is dangerous because users may reasonably expect a chat-only research interaction, while the skill silently persists intermediate outputs and task metadata to disk.