Back to skill

Security audit

OpenXSys

Security checks for vulnerabilities and agentic risk

Overview

ClawScan marks this skill suspicious because it reports persistent learning assessments without a clear privacy notice and many broad activation conditions, though some other alerts appear to match ordinary explanatory text.

Consider the suspicious verdict and the privacy concern before installing, especially if Learn mode may be used with personal learning information. The scan is incomplete: it covered 99.3% of components, left one component partially inspected, could not resolve some references, excluded binary images from some checks, and reported one degraded analyzer. The findings support caution, but do not establish malicious intent.

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
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Output HandlingUnvalidated Output Injection, Cross-Context Output, Unbounded Output
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
Findings (46)

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 1817)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">System</span><span class="depth-name">What this run leaves behind</span></div>

Invocation examples overlap with everyday requests without defining when OpenXSys should not activate.

Medium
Category
Trigger Abuse
Confidence
94% confidence
Finding

The invocation table includes ordinary requests such as reviewing material, choosing between directions, and “Continue.” The README does not explicitly restrict these phrases to an already-open OpenXSys session or provide exclusions, making the activation boundary unclear.

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 describes broad activities—thinking, making, connecting, and growing—while the policy permits implicit invocation. Although the default prompt explicitly names the skill, the manifest does not establish which ordinary requests qualify for implicit activation or which should be excluded.

Content

No source excerpt is available for this finding.

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
75% confidence
Finding

Skill enables autonomous high-impact decisions without human-in-the-loop verification. Critical operations (destructive commands, financial transactions, data deletion) should require explicit user confirmation.

Content

Scanner excerpt · capabilities/methods/1-see/estimate.md (reported line 35)May include surrounding context.

md
- **One path only**, so there is nothing to catch an error.
- **Hidden assumptions** inside a factor that looked like a fact.
- **Units and time periods mixed**: per year against per month, stock against flow.
- **Estimating what should have been looked up**, or quoting a figure without checking that it is plausible.

## It works with

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
75% confidence
Finding

Skill enables autonomous high-impact decisions without human-in-the-loop verification. Critical operations (destructive commands, financial transactions, data deletion) should require explicit user confirmation.

Content

Scanner excerpt · capabilities/methods/2-explain/bottleneck.md (reported line 35)May include surrounding context.

md
- **Ten bottlenecks.** A list of everything that is slow is not a finding.
- **Local efficiency.** Making a non-constraint faster only builds a bigger pile.
- **No goal**, so anything can be called a limit.
- **Assuming the limit is a resource** ("we need more people") without checking whether a rule is the limit.
- **Naming the constraint and stopping**, with no word on how to use it better before spending on it.

## It works with

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
75% confidence
Finding

Skill enables autonomous high-impact decisions without human-in-the-loop verification. Critical operations (destructive commands, financial transactions, data deletion) should require explicit user confirmation.

Content

Scanner excerpt · capabilities/methods/4-focus/vital_few.md (reported line 34)May include surrounding context.

md
- **Quoting 80/20 with no data.**
- **Categories that create the concentration.** Grouping things until one group is large proves nothing.
- **The wrong measure**: ranking by what is easy to count, not by what matters.
- **Cutting the many** without asking what they support.
- **A single period**, taken as permanent.

## It works with

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
75% confidence
Finding

Skill enables autonomous high-impact decisions without human-in-the-loop verification. Critical operations (destructive commands, financial transactions, data deletion) should require explicit user confirmation.

Content

Scanner excerpt · capabilities/methods/lenses/bound.md (reported line 30)May include surrounding context.

md
## Where it goes wrong

- Widening until the system is "everything" and nothing can be said.
- Listing stakeholders without asking what each one changes.
- Treating the first boundary as a fact given by the world.

## It works with

Learn mode persists personal learning assessments without requiring a user-facing privacy notice.

Medium
Category
Output Handling
Confidence
87% confidence
Finding

Learn mode explicitly requires persisting predictions and judgments about what the person understands, while its activation language only describes a teaching interaction. Without an existing informed consent covering these records, this can create unexpected personal profiling and expose potentially sensitive learning assessments to later agents or workspace readers; the general statement about user ownership does not itself provide that notice.

Content

No source excerpt is available for this finding.

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
75% confidence
Finding

Skill enables autonomous high-impact decisions without human-in-the-loop verification. Critical operations (destructive commands, financial transactions, data deletion) should require explicit user confirmation.

Content

Scanner excerpt · core/relation.md (reported line 25)May include surrounding context.

md
## Gaps and bridge systems

A gap exists when a difference, understanding, or capability held by one system cannot produce a usable effect in another. Typical causes include invisibility, incompatible meaning, mismatched timing, missing authority, no execution path, no confirmation, or feedback that cannot reach the party able to change the next action.

When bridging a gap, first ask whether observation, translation, action, coordination, or feedback is missing. Then organize a bridge system. It may contain a natural-language agreement, people, AI, existing tools, generated code, shared data, or a new division of work. The bridge has its own purpose, state, boundary, capabilities, and failure conditions; it can be replaced or continue to evolve.

Subjective, everyday events can activate the evolution engine.

Medium
Category
Trigger Abuse
Confidence
92% confidence
Finding

L03 makes 'something is worth keeping' and a finished deliverable sufficient reading conditions. These conditions do not distinguish an ordinary conversational outcome from an authorized update to a maintained system.

Content

No source excerpt is available for this finding.

Common conversational phrases and session events are overly broad triggers.

Medium
Category
Trigger Abuse
Confidence
96% confidence
Finding

L03 includes an opening, a returning person, and the phrase 'and then?' as reasons to activate this engine. Those events can occur in almost any conversation, including simple follow-up questions unrelated to proactive exploration.

Content

No source excerpt is available for this finding.

Needing facts or an explanation is too broad an activation condition.

Medium
Category
Trigger Abuse
Confidence
95% confidence
Finding

L003 activates the engine whenever facts are needed or something must be explained, which includes routine factual questions. L004 describes backstage execution beneath methods but does not constrain activation to a selected research or modelling task.

Content

No source excerpt is available for this finding.

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
75% confidence
Finding

Skill enables autonomous high-impact decisions without human-in-the-loop verification. Critical operations (destructive commands, financial transactions, data deletion) should require explicit user confirmation.

Content

Scanner excerpt · experience/display/forms.md (reported line 56)May include surrounding context.

md
## Choosing among tiers at full power

When a video, a concept animation or a deck would serve the need and its dependency is ready, make it without asking. This is what full power means for media. Judge "would serve the need" honestly:

- **A short video** serves when the reveal is a movement or a sequence, or when the person makes content for an audience that watches rather than reads.
- **A concept animation** serves when the idea is mathematical or physical and static figures would need paragraphs to explain.

Everyday speech and vague discomfort are presented as sufficient conditions to begin a run.

Medium
Category
Trigger Abuse
Confidence
94% confidence
Finding

The AI-host guidance says ordinary speech, files, and vague discomfort can all begin a run, and treats an everyday statement about not knowing what to ask as an entry point. It does not distinguish a request to invoke this skill from ordinary conversation or requests intended for other skills.

Content

No source excerpt is available for this finding.

The activation-routing condition covers essentially any first mention of a subject.

Medium
Category
Trigger Abuse
Confidence
93% confidence
Finding

The table routes an ordinary expression pointing to a subject for the first time to activation. Without a requirement that the user request system creation or already be working within an activated system, this condition overlaps with most new conversational topics.

Content

No source excerpt is available for this finding.

Activation accepts broad conversational and environmental signals without a clear invocation boundary.

Medium
Category
Trigger Abuse
Confidence
98% confidence
Finding

Entry points include an ordinary sentence, curiosity, a wish, a correction, “continue,” and changes in files or the environment. Host delivery establishes that an event exists, but does not specify which events or conversational expressions should activate this skill rather than remain ordinary input.

Content

No source excerpt is available for this finding.

Run-start conditions do not distinguish ordinary conversation from an intentional system invocation.

Medium
Category
Trigger Abuse
Confidence
86% confidence
Finding

The architecture says a conversation or an event observed by the host can start a run, without identifying the conversational intent or event filter required. Although scheduled entries are explicitly authorized, the conversation and observed-event cases lack an equally clear activation boundary.

Content

No source excerpt is available for this finding.

The trigger “where am I now?” can activate a system panorama for unrelated location or progress questions.

Medium
Category
Trigger Abuse
Confidence
94% confidence
Finding

The read condition treats an everyday question as a direct trigger without requiring an active OpenXSys context. The same phrase could refer to geographic location, navigation, or progress within an unrelated task.

Content

No source excerpt is available for this finding.

The unqualified trigger “continue” can invoke system-return behavior when the user only wants to resume an unrelated task.

Medium
Category
Trigger Abuse
Confidence
95% confidence
Finding

The activation condition lists “continue” without requiring that it refer to an OpenXSys system; line 46 reinforces that invocation. The later guidance resolves competing system targets but does not exclude ordinary requests to continue a response or unrelated workflow.

Content

No source excerpt is available for this finding.

The invocation example “Where does the time go?” overlaps with ordinary conversation.

Low
Category
Trigger Abuse
Confidence
92% confidence
Finding

This phrase can be a casual observation rather than a request to analyze stocks and flows. Listing it under “Also heard as” without an activation constraint risks selecting the method when no flow analysis was intended.

Content

No source excerpt is available for this finding.

Broad recurring-problem questions are presented as invocation phrases without a scope constraint.

Low
Category
Trigger Abuse
Confidence
87% confidence
Finding

Questions such as “Why can't we get past ___?” can request ordinary troubleshooting or personal advice rather than feedback-loop analysis. The invocation examples do not distinguish those requests from intentional use of this method.

Content

No source excerpt is available for this finding.

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

Low
Category
Trigger Abuse
Confidence
97% confidence
Finding

This wording is common in general explanations and tutoring across virtually any topic. It does not establish that the user wants the system-level decomposition required by Open.

Content

No source excerpt is available for this finding.

Generic questions about what is happening are treated as X-ray invocation phrases.

Low
Category
Trigger Abuse
Confidence
94% confidence
Finding

“What is actually going on with ___?” and “What is behind ___?” can request a status update or straightforward factual explanation. The examples do not require a recurring pattern or a request for structural causal analysis.

Content

No source excerpt is available for this finding.

Everyday troubleshooting questions can activate the Bottleneck method too broadly.

Low
Category
Trigger Abuse
Confidence
90% confidence
Finding

“Why can't we go faster?” and “Where is ___ stuck?” do not necessarily ask for identification of a system-wide throughput constraint. They can refer to a local delay or an ordinary progress check.

Content

No source excerpt is available for this finding.

“Does it have to be this way?” is an ambiguous, everyday invocation phrase.

Low
Category
Trigger Abuse
Confidence
95% confidence
Finding

This question can express dissatisfaction or ask about a simple alternative without requesting first-principles decomposition. Other listed alternatives, such as asking what something is made of, can likewise be answered factually without a rebuild.

Content

No source excerpt is available for this finding.