Back to skill

Security audit

OpenXSys

Security checks for vulnerabilities and agentic risk

Overview

ClawScan marks this skill suspicious with medium confidence: the findings raise concerns about broad activation, persistent records and system changes, and a reported mismatch between its stated purpose and an included order-processing example.

Consider the suspicious verdict and HIGH risk assessment recommendation not to install. If you are still considering it, review the broad activation rules and the reported retention and system-change practices first. The scan was partial: it covered 99.3% of components, with one analyzer degraded and some references unresolved, so the evidence is incomplete.

Vulnerability Patterns
  • Skill Instruction HijackingAlters the agent's session goals or safety constraints when the skill loads
  • Agent Memory PoisoningWrites attacker-controlled rules into memory that affect later sessions
  • Remote Payload Retrieval and ExecutionFetches external code whose behavior can change after review
  • Embedded Malicious CodeShips malicious scripts inside the skill and executes them locally
  • Unauthorized Access and Privilege EscalationObtains permissions beyond the task's legitimate needs
Vulnerability Patterns
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Output HandlingUnvalidated Output Injection, Cross-Context Output, Unbounded Output
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
Findings (36)

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
92% confidence
Finding

The supplied code implements order fulfillment rather than the declared systems-analysis and continuity functions. Its inventory mutations and notification calls are not described capabilities. The path suggests it may be a playground sample, but the supplied chunk alone does not establish how it supports the declared skill. No specific filesystem or network access can be established because inventory and notifier implementations are not included.

Content

No source excerpt is available for this finding.

Hidden Instructions

High
Category
Prompt Injection
Confidence
70% confidence
Finding

Hidden instructions were detected in comments or invisible text. These could contain malicious directives. Manual review is recommended.

Content

Scanner excerpt · experience/playground/replays/industry/index.html (reported line 1818)May include surrounding context.

html
</div>
</section>

<!-- ===== 16 SYSTEM ===== -->
<section class="section block-paper" id="s16" data-depth="8">
  <div class="wrap">
    <div class="depth-tag"><span class="depth-num">系统</span><span class="depth-name">这次运行留下了什么</span></div>

Activation examples overlap with ordinary requests without defining when the skill should not activate.

Medium
Category
Trigger Abuse
Confidence
94% confidence
Finding

The README says users can speak normally and presents generic requests such as reviewing materials, choosing between directions, and continuing as entry points. It does not distinguish these requests from ordinary host tasks or require an explicitly active OpenXSys system, creating potential for unintended invocation.

Content

No source excerpt is available for this finding.

Implicit invocation is enabled without a specific activation scope in the manifest.

Medium
Category
Trigger Abuse
Confidence
85% confidence
Finding

The short description covers broad activities such as understanding, thinking, producing results, and connecting, while the policy enables implicit invocation. This manifest supplies neither concrete implicit activation conditions nor exclusions separating those activities from everyday assistant work.

Content

No source excerpt is available for this finding.

“Help me really understand ___” is an overly broad alternative invocation.

Medium
Category
Trigger Abuse
Confidence
96% confidence
Finding

“帮我真正搞懂 ___” can introduce almost any educational request, including requests that need only a definition or short explanation. It does not establish that the user wants the skill's whole-system decomposition.

Content

No source excerpt is available for this finding.

Generic “what is going on?” questions are treated as invocations for structural causal analysis.

Medium
Category
Trigger Abuse
Confidence
95% confidence
Finding

“___ 到底是怎么回事?” and “___ 背后是什么?” are common requests for clarification or factual context. Although the method later focuses on recurring events, the invocation alternatives do not require recurrence or an intent to investigate underlying structures.

Content

No source excerpt is available for this finding.

The invocation examples include generic requests for advice without limiting activation to best-practice selection.

Medium
Category
Trigger Abuse
Confidence
96% confidence
Finding

L05 includes “这件事我该怎么做?” (“How should I do this?”), which can occur in almost any assistance request. The file does not distinguish this everyday wording from an explicit request to research, compare, and adapt established practices.

Content

No source excerpt is available for this finding.

Short everyday choice questions are presented as invocation phrases without scope constraints.

Medium
Category
Trigger Abuse
Confidence
93% confidence
Finding

“选哪个?” (“Which one?”) and “值不值?” (“Is it worth it?”) can refer to trivial selections or conversational follow-ups. No activation condition distinguishes those cases from requests for the skill's structured decision analysis.

Content

No source excerpt is available for this finding.

A generic configuration request can activate a much broader capability-building workflow.

Medium
Category
Trigger Abuse
Confidence
88% confidence
Finding

“帮我把这个配起来。” (“Help me configure this.”) need not mean building a reusable capability and validating it on real work. The activation examples do not require that broader intent or exclude one-off configuration help.

Content

No source excerpt is available for this finding.

Generic organization and quality questions are treated as design-skill invocation examples.

Medium
Category
Trigger Abuse
Confidence
88% confidence
Finding

“这个该怎么组织?” (“How should this be organized?”) and “这东西做得好应该是什么样?” (“What should this look like when done well?”) can request lightweight advice rather than a new system, process, or product design. The file supplies no exclusion or requirement that the user want a multi-option design exercise.

Content

No source excerpt is available for this finding.

The phrase “write this out” does not distinguish ordinary drafting from finished-artifact production.

Medium
Category
Trigger Abuse
Confidence
92% confidence
Finding

“把这个写出来。” (“Write this out.”) is a common request that may only ask for a short response or draft. Here it is an invocation example for producing a completed artifact, without a condition requiring a standalone deliverable.

Content

No source excerpt is available for this finding.

Conversational reflection questions can trigger a formal retrospective without an explicit retrospective intent.

Medium
Category
Trigger Abuse
Confidence
91% confidence
Finding

“我们学到了什么?” (“What did we learn?”) and “那样做管用了吗?” (“Did that work?”) can be requests for a summary or status check. The invocation guidance does not distinguish those requests from a retrospective that compares intentions and outcomes and changes system records.

Content

No source excerpt is available for this finding.

Advice about what to track is not distinguished from a request to establish persistent monitoring.

Medium
Category
Trigger Abuse
Confidence
92% confidence
Finding

The examples include advisory questions such as what to track, while the stated output requires saving monitoring conditions and checking them on later returns. This conflates advice with establishing persistent state and future activity; the explicit authorization requirement for out-of-session scheduling mitigates, but does not resolve, that narrower scope issue.

Content

No source excerpt is available for this finding.

Route-selection examples overlap with ordinary explanatory, decision-making, and learning requests without a clear activation boundary.

Medium
Category
Trigger Abuse
Confidence
80% confidence
Finding

Examples such as “为什么会发生 ___?” and “我们该 ___ 还是 ___?” encompass common everyday requests and can select a multi-stage systems workflow without an explicit request for that workflow. Although later instructions contextualize route selection, they do not clearly distinguish when ordinary conversation should remain outside route activation.

Content

No source excerpt is available for this finding.

The learning workflow requires persistent records of the user's predictions and demonstrated understanding without requiring notice of that profiling.

Medium
Category
Output Handling
Confidence
87% confidence
Finding

Learning mode requires persisting the user's predictions, demonstrated understanding, and remaining difficulties, without specifically disclosing that retention or checking consent for it. Opting into instruction is not necessarily opting into a durable learner profile, which can expose personal information to later sessions or other readers of the system files.

Content

No source excerpt is available for this finding.

The workflow directs deletion of supposedly obsolete guidance without warning about loss of existing instructions.

Medium
Category
Output Handling
Confidence
88% confidence
Finding

L33 instructs the agent to remove guidance it judges unnecessary, which can permanently change system behavior. Although the file calls for briefly announcing writes and preserving user edits, it does not warn about deletion-related information loss or require a recoverable copy before removing guidance.

Content

No source excerpt is available for this finding.

Activation includes ordinary conversational follow-ups and loosely defined session events.

Medium
Category
Trigger Abuse
Confidence
94% confidence
Finding

The activation description includes ‘然后呢’ (‘then what?’), returning, and starting, without limiting these signals to an established exploration task. These conditions overlap with routine conversation and could trigger unsolicited exploration when the user simply wants a direct continuation.

Content

No source excerpt is available for this finding.

Generic requests such as ‘write it out’ and ‘give me a plan’ are treated as activation signals.

Medium
Category
Trigger Abuse
Confidence
91% confidence
Finding

The invocation examples include ‘写出来’ (‘write it out’) and ‘给我一个方案’ (‘give me a plan’), while another condition is merely that it is time to do something. The file does not require a specific research-artifact context or distinguish these requests from ordinary drafting and planning.

Content

No source excerpt is available for this finding.

The activation conditions encompass almost any factual or explanatory request.

Medium
Category
Trigger Abuse
Confidence
95% confidence
Finding

‘需要事实’ (‘need facts’) and explaining ‘why’ are broad conditions common to everyday assistant use. No activation boundary distinguishes a substantive investigation from a simple factual answer, so routine questions could invoke the full modeling and record-maintenance workflow.

Content

No source excerpt is available for this finding.

Activation is triggered by any ordinary expression that first refers to an object, without a clear intent or scope requirement.

Medium
Category
Trigger Abuse
Confidence
90% confidence
Finding

The activation row routes an ordinary expression that first points to an object into the activation workflow. This can encompass routine questions or incidental mentions, and the file provides no exclusions or requirement that the user intend to start an OpenXSys system.

Content

No source excerpt is available for this finding.

Activation includes everyday expressions such as “continue” and broadly defined events without clear invocation boundaries.

Medium
Category
Trigger Abuse
Confidence
96% confidence
Finding

The entry conditions include ordinary curiosity, concerns, wishes, corrections, and “继续” (“continue”), as well as file changes and tool results. No explicit invocation requirement or exclusions distinguish a request to activate this system from ordinary conversation or routine events.

Content

No source excerpt is available for this finding.

The workflow silently inventories local capabilities and persists the results without requiring disclosure of that persistence.

Medium
Category
Output Handling
Confidence
87% confidence
Finding

The workflow directs quiet enumeration of user-level skill folders and available integrations, followed by persistence into a root system for future sessions, without clearly establishing consent for that inventory's scope and retention. Tool names, account references, and capability metadata can reveal sensitive organizational or personal context even without credentials; the existing non-secret and authorization safeguards reduce the risk but do not clearly constrain this default cross-session inventory.

Content

No source excerpt is available for this finding.

Everyday questions about money or time are listed as invocation alternatives without a systems-analysis constraint.

Low
Category
Trigger Abuse
Confidence
88% confidence
Finding

“钱到底是从哪里来的?” and “时间都去哪了?” can be ordinary factual questions or conversational remarks. Listing them as alternative invocations without requiring an intent to trace stocks and flows could activate this method unnecessarily.

Content

No source excerpt is available for this finding.

Generic growth and repair questions can trigger feedback-loop analysis without establishing that it is wanted.

Low
Category
Trigger Abuse
Confidence
87% confidence
Finding

“___ 为什么一直在涨?” and “为什么每次修好了又坏?” overlap with ordinary factual questions and troubleshooting requests. The invocation examples do not distinguish those requests from a request to investigate systemic feedback.

Content

No source excerpt is available for this finding.

Broad questions about being slow or stuck do not clearly delimit bottleneck-method activation.

Low
Category
Trigger Abuse
Confidence
86% confidence
Finding

“我们为什么快不起来?” and “___ 卡在哪里?” can describe a local technical fault or a request for a status update. The invocation text does not require a measurable system goal or an intent to identify the constraint on overall output.

Content

No source excerpt is available for this finding.