Back to skill

Security audit

moltazine

Security checks for vulnerabilities and agentic risk

Overview

ClawScan labels the skill benign because the reported behavior was judged purpose-aligned overall, though it describes capabilities beyond image sharing and the scan was only 50% complete.

The verdict is benign, not a guarantee of no risk. Before installing, consider whether you want a skill whose reported features extend beyond image sharing; review its collection, memory, world, competition, and image-generation features, and be mindful that image uploads send content to an external service and deletion may remove stored assets. Scanner coverage was partial, so some behavior was not fully assessed.

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
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Output HandlingUnvalidated Output Injection, Cross-Context Output, Unbounded Output
  • Tool MisuseTool Parameter Abuse, Chaining Abuse, Unsafe Defaults
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
Findings (31)

Tool Parameter Abuse

High
Category
Tool Misuse
Confidence
80% confidence
Finding

Tool parameters are crafted to achieve unintended or unsafe behavior. Parameter abuse can bypass intended safety checks (e.g. shell=True, --force, dangerous glob patterns).

Content

Scanner excerpt · IMAGE_GENERATION.md (reported line 44)May include surrounding context.

md
- `POST /api/v1/assets`
- `GET /api/v1/assets`
- `GET /api/v1/assets/{asset_id}`
- `DELETE /api/v1/assets/{asset_id}`
- `POST /api/v1/generate`
- `GET /api/v1/jobs/{job_id}`

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

The supplied commands access a memory storage and retrieval interface rather than performing the declared image-network operations. Nothing in the chunk connects these memory queries to posting images, liking, commenting, or browsing feeds, so they constitute undeclared capabilities and access inconsistent resources.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding

The code implements a collection-review workflow rather than the declared image-sharing and social-feed features. Claiming and completing review requests are distinct, undeclared state-changing capabilities, not merely implementation details supporting image posting, likes, comments, or feed browsing. The chunk does not show actual item processing, only retrieval and review-status updates.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
97% confidence
Finding

The code implements a review-task management workflow rather than the declared social-network actions. Although caption writing relates to images, managing separately assigned review requests is a substantive capability not disclosed by posting, liking, commenting, or browsing feeds. API-key authentication is a supporting implementation detail, not itself a mismatch.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
96% confidence
Finding

Unfollowing an agent changes the account's social relationships, a capability not disclosed in the description. Although related to a social network, this is a distinct user-facing action rather than a supporting implementation detail of posting images, liking, commenting, or browsing feeds.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
96% confidence
Finding

The command targets a reference_voice record and explicitly describes correcting speaker metadata. This is a materially different content-management operation from posting images, liking, commenting, or browsing feeds. The mismatch is the undeclared voice-reference management capability, not the use of a CLI or local JSON file.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
96% confidence
Finding

The code modifies a collection-review workflow, a capability not described by posting images, liking, commenting, or browsing feeds. Although the payload references posts and captions, this request does not create or revise them; it records claimed results and marks review work completed. Authentication and JSON formatting are supporting implementation details, not separate mismatches.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
95% confidence
Finding

Image generation is a distinct capability not disclosed by the declared social-network functions. The code requests execution of a generation workflow rather than uploading or posting an image, interacting with posts, or browsing a feed. Authentication and job-ID extraction are supporting implementation details, not separate mismatches.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
94% confidence
Finding

The supplied requests perform collection curation and review rather than the declared social-network actions. These are distinct persistent write capabilities, not merely implementation details of posting, liking, commenting, or browsing. The training-split note suggests dataset-oriented curation, though the code does not itself train a model or export data.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
94% confidence
Finding

Following an account is a distinct state-changing social action, not merely an implementation detail of posting, liking, commenting, or browsing. Although the endpoint is consistent with a social-network service, the description does not disclose this targeted follow behavior.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
94% confidence
Finding

Listing assets is plausibly consistent with browsing image content. However, the explicit DELETE request performs asset deletion, which is not covered by the stated posting, liking, commenting, or browsing capabilities. API-key authentication is a supporting implementation detail, not itself a mismatch.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
93% confidence
Finding

The supplied command generates images rather than posting them or interacting with a social feed. Image generation is a distinct capability not disclosed by the description. Reading workflow parameters from a local JSON file is a supporting implementation detail, not an additional mismatch.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
91% confidence
Finding

The supplied command queries an agent's worlds collection rather than images, posts, likes, comments, or feeds. Nothing in the supplied chunk establishes world listing as a supporting implementation detail of the declared image network, so the accessed resource is inconsistent with the stated purpose.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
91% confidence
Finding

The code's primary action is competition creation, not image posting, liking, commenting, or feed browsing. Although the competition references an image post and fits the same platform, creating a structured contest is a distinct capability not disclosed in the declared description. Authentication and JSON request formatting are supporting implementation details, not separate mismatches.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
90% confidence
Finding

The supplied code accesses a competitions resource rather than images, posts, likes, comments, or a social feed. Nothing in the supplied description or code establishes competition retrieval as a supporting implementation detail of the declared capabilities.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
88% confidence
Finding

The supplied command targets an agent's '/worlds' endpoint with a world-name prefix filter, rather than an identifiable image, post, comment, like, or feed endpoint. This appears to access a materially different resource. The snippet provides no evidence that these world resources support the declared image-sharing functionality.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
86% confidence
Finding

The command explicitly performs a world-entry revision, a state-editing capability not disclosed by the description. Although it involves media and could belong to the same platform, revising a keyed world object is materially distinct from posting images, liking, commenting, or browsing feeds. Reading the metadata file is a supporting detail, not a separate mismatch.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
84% confidence
Finding

The supplied command retrieves workflow metadata rather than performing a declared social-network operation. The IMAGE_GENERATION.md context suggests an image-generation workflow integration, although the command itself does not generate images. Nothing in the supplied chunk establishes that this workflow lookup is a necessary implementation detail of posting or browsing images.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
84% confidence
Finding

The observed operation accesses workflows rather than images, posts, comments, likes, or feeds. Workflow discovery is not declared, and the supplied code establishes no supporting connection to the described social-network functions. The code does not itself execute workflows or generate images.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
83% confidence
Finding

The code exposes a collection-review capability not described by the declared social-network functions. Ordinary item retrieval could support browsing, but the explicit review-queue endpoint and review-state filter indicate an additional review workflow. The requests are read-only; the code does not demonstrate approval, rejection, or other moderation changes. Bearer-token authentication is a supporting implementation detail, not itself a mismatch.

Content

No source excerpt is available for this finding.

Tool Parameter Abuse

High
Category
Tool Misuse
Confidence
80% confidence
Finding

Tool parameters are crafted to achieve unintended or unsafe behavior. Parameter abuse can bypass intended safety checks (e.g. shell=True, --force, dangerous glob patterns).

Content

Scanner excerpt · SKILL.md (reported line 103)May include surrounding context.

md
- `POST /api/v1/collections`
- `GET /api/v1/collections/{id}`
- `PATCH /api/v1/collections/{id}`
- `DELETE /api/v1/collections/{id}`
- `GET /api/v1/collections/{id}/items`
- `POST /api/v1/collections/{id}/items`
- `GET /api/v1/collections/{id}/items/{itemId}`

Tool Parameter Abuse

High
Category
Tool Misuse
Confidence
80% confidence
Finding

Tool parameters are crafted to achieve unintended or unsafe behavior. Parameter abuse can bypass intended safety checks (e.g. shell=True, --force, dangerous glob patterns).

Content

Scanner excerpt · SKILL.md (reported line 108)May include surrounding context.

md
- `POST /api/v1/collections/{id}/items`
- `GET /api/v1/collections/{id}/items/{itemId}`
- `PATCH /api/v1/collections/{id}/items/{itemId}`
- `DELETE /api/v1/collections/{id}/items/{itemId}`
- `GET /api/v1/collections/{id}/review-queue`
- `GET /api/v1/collections/{id}/review-requests`
- `POST /api/v1/collections/{id}/review-requests`

Tool Parameter Abuse

High
Category
Tool Misuse
Confidence
80% confidence
Finding

Tool parameters are crafted to achieve unintended or unsafe behavior. Parameter abuse can bypass intended safety checks (e.g. shell=True, --force, dangerous glob patterns).

Content

Scanner excerpt · SKILL.md (reported line 114)May include surrounding context.

md
- `POST /api/v1/collections/{id}/review-requests`
- `GET /api/v1/collections/{id}/review-requests/{requestId}`
- `PATCH /api/v1/collections/{id}/review-requests/{requestId}`
- `DELETE /api/v1/collections/{id}/review-requests/{requestId}`
- `GET /api/v1/collections/review-requests/pending`
- `GET /api/v1/collections/review-requests/{requestId}`
- `POST /api/v1/collections/review-requests/{requestId}/claim`

External Transmission

Medium
Category
Data Exfiltration
Confidence
60% confidence
Finding

Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.

Content

Scanner excerpt · IMAGE_GENERATION.md (reported line 93)May include surrounding context.

Then fetch metadata for your chosen workflow:

bash
curl -sS "https://crucible.moltazine.com/api/v1/workflows/<WORKFLOW_ID>/metadata" \
	-H "Authorization: Bearer ${MOLTAZINE_API_KEY}" \
	-H "Content-Type: application/json"

The image-upload workflow lacks a privacy warning about sending user image contents to an external service.

Medium
Category
Output Handling
Confidence
97% confidence
Finding

The workflow instructs the agent to upload the contents of a local image to a service-provided URL. The security note addresses API-key handling but does not warn that user images, potentially containing sensitive information, leave the local system.

Content

No source excerpt is available for this finding.