Back to skill

Security audit

huawei-cloud-cloudrobo-model-workflow

Security checks for vulnerabilities and agentic risk

Overview

This workflow needs review because its recovery instructions recommend repeating cancelled robot tasks without checking whether physical conditions remain safe.

Before use, change recovery handling so cancellation stops the workflow until the operator requests a fresh run. Require confirmation that the robot has stopped and its work area is safe before dispatch or retry. Preserve the existing approval requirement for cloud writes, verify the dataset directory and destination before upload, and protect cloud keys and exported robot certificates. This review covers the bundled instructions, not the external CLI implementation or physical safety systems.

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
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Output HandlingUnvalidated Output Injection, Cross-Context Output, Unbounded Output
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
Findings (9)

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
88% confidence
Finding

The supplied chunk performs dataset catalog writes and release-status changes, not training, deployment, evaluation, or read-only asset queries. Although dataset registration could support a training workflow, creating and releasing assets are substantive lifecycle-management operations not disclosed by the declared asset-query capability. The chunk references an existing dataset path; it does not demonstrate uploading dataset contents.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
88% confidence
Finding

The supplied chunk performs dataset ingestion and asset lifecycle management, neither of which is declared in the description. These are substantive asset-writing operations, not just implementation details of asset querying. Although they could support a training workflow, the chunk does not establish that connection or perform training, deployment, or evaluation.

Content

No source excerpt is available for this finding.

Credential Access

High
Category
Privilege Escalation
Confidence
60% confidence
Finding

Code accesses credential files (SSH keys, AWS credentials, etc.). This could indicate credential theft attempts.

Content

Scanner excerpt · references/cli-installation-guide.md (reported line 33)May include surrounding context.

bash
# Set required environment variables
# ⚠️ NEVER commit real keys. Use a secrets manager or .env file.
export HUAWEI_CLOUD_AK="your-access-key-id"
export HUAWEI_CLOUD_SK="your-secret-access-key"

The workflow executes tasks on a real robot and retries failures without warning the user about physical actuation or repeated execution.

High
Category
Output Handling
Confidence
95% confidence
Finding

The diagram routes failed or cancelled evaluations back to task creation, which immediately actuates a physical robot, without a renewed authorization or physical-readiness check. Initial robot selection does not establish that a partially completed task is safe to repeat, and retrying a cancelled task can defeat an operator's stop decision.

Content

No source excerpt is available for this finding.

Real-robot execution lacks a physical-safety warning before dispatch.

High
Category
Output Handling
Confidence
96% confidence
Finding

The standalone template proceeds from querying online robots directly to an immediately executing dispatch, without an operator readiness gate or physical-safety checks. Online status and execution-duration limits do not establish safe motion, particularly for a newly trained manipulation policy.

Content

No source excerpt is available for this finding.

The local-dataset workflow uploads user files to cloud storage without a privacy warning.

Medium
Category
Output Handling
Confidence
88% confidence
Finding

Case C instructs the agent to import a local directory into OBS and release the resulting dataset version. Although the upload is described elsewhere, the workflow provides no user-facing warning that local dataset contents will leave the machine and be stored in Huawei Cloud, where sensitive data may require additional handling.

Content

No source excerpt is available for this finding.

The PowerShell workaround executes a cloud training-task creation command without warning about its side effects.

Medium
Category
Output Handling
Confidence
90% confidence
Finding

The recommended subprocess example runs cloudrobo train create-task with the contents of config.json, rather than merely demonstrating JSON parsing. It does not warn that running the example submits that configuration to CloudRobo and creates a training task in the active workspace.

Content

No source excerpt is available for this finding.

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
97% confidence
Finding

The recovery guide explicitly recommends recreating a manually cancelled task and says that retry creation auto-executes. This can override a deliberate stop and repeat physical motion after the environment or robot state has changed, risking injury or equipment damage.

Content

Scanner excerpt · references/fault-recovery.md (reported line 68)May include surrounding context.

md
| Task `FAILED` | Model inference error or robot execution exception | Check `show-task-result` logs |
| Task `CANCELLED` | Manually cancelled | Recreate task |

> After fixing, create a new task to retry (`create-task` auto-executes).

Local dataset import lacks a warning about transferring user data to the cloud service.

Medium
Category
Output Handling
Confidence
94% confidence
Finding

The workflow instructs the agent to import a local dataset into CloudRobo and release its version. It does not warn that local dataset contents will leave the user's machine or ask the user to check for sensitive data before upload.

Content

No source excerpt is available for this finding.