Back to skill

Security audit

huawei-cloud-cloudrobo-r2c

Security checks for vulnerabilities and agentic risk

Overview

This robot-control skill needs review because its custom adapter can move hardware during testing advertised as safe from movement.

Use a dummy adapter until the motion-safety issue is corrected. Dry-run currently cannot guarantee that startup or shutdown will avoid movement. Require explicit authorization before real hardware execution. Replace shared temporary paths with a private temporary directory and use an external timeout for tests. Review the cloud destination and camera mappings because dry-run still transmits observations. Protect the credential bundle and verify external packages; this review inspected all bundled files but did not execute or validate the external SDK.

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
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Output HandlingUnvalidated Output Injection, Cross-Context Output, Unbounded Output
  • Tool MisuseTool Parameter Abuse, Chaining Abuse, Unsafe Defaults
  • MCP Least PrivilegeUnderdeclared Capability, Wildcard Permission, Missing Permission Declaration
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
Findings (10)

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

Both commands perform operations explicitly excluded by the declared description and delegated to the robot skill. Although registration and credential export can be prerequisites for running the client, this chunk directly instructs those operations rather than using an existing credential bundle to start the client.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
95% confidence
Finding

The supplied executable action is a CLI test invocation, which has a materially different apparent purpose from running the R2C client. No client startup, credential loading, hardware adapter initialization, or control-loop execution is shown. Although testing could support client development, this chunk presents testing as its sole action rather than as a supporting step. The unavailable script contents limit conclusions about any additional behavior.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
90% confidence
Finding

The supplied example explicitly starts a cloud adapter and replaces its default policy client with custom model inference. This is a different operational role from the declared robot-side client and hardware-adapter functionality. The generic mention of a custom adapter does not clearly disclose cloud-side inference hosting. The chunk does not demonstrate robot registration, certificate export, or malicious behavior.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
84% confidence
Finding

The supplied chunk concerns cloud IAM credentials, a distinct authentication mechanism not described as part of the robot-side client. No connection to the declared client functionality is shown. The values are placeholders, so the code does not itself expose real credentials, access cloud resources, or perform IAM operations.

Content

No source excerpt is available for this finding.

Incomplete referenced artifact analysis

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 290)May include surrounding context.

md
- [Client Config Reference](references/client-config-catalog.md) — client_config.yaml field mapping, Zenoh QoS settings

The custom adapter moves the robot during connection, contradicting the guide's dry-run safety assurance.

High
Category
Excessive Agency
Confidence
99% confidence
Finding

The custom adapter calls move_joints(initial) inside connect(), which runs during startup independently of the control loop's suppression of received actions. Because the example configuration supplies initial joint positions, following the advertised dry-run workflow can still cause unexpected physical movement, risking injury or equipment damage.

Content

No source excerpt is available for this finding.

Tool Parameter Abuse

High
Category
Tool Misuse
Confidence
95% confidence
Finding

Although rm -f is ordinary test cleanup, the surrounding workflow records to a predictable shared /tmp pathname and deletes it without establishing ownership or creating it exclusively. This can overwrite or delete an existing recording and permits local interference with the output path; rm itself does not follow symlinks to delete their targets.

Content

Scanner excerpt · scripts/test-cli-commands.sh (reported line 101)May include surrounding context.

sh
--log-level INFO 2>&1 || true
    if [ -f /tmp/r2c_test_observations.pkl ]; then
        echo "  Recording file created: $(ls -la /tmp/r2c_test_observations.pkl)"
        rm -f /tmp/r2c_test_observations.pkl
    else
        echo "  Warning: recording file not created"
    fi

Skill declares no tool scope ('permissions' or 'allowed-tools') but code capabilities were detected: file_read, file_write.

Medium
Category
MCP Least Privilege
Confidence
70% confidence
Finding

Without declared permissions the skill's intent is opaque and cannot be validated.

Content

No source excerpt is available for this finding.

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
85% 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 · references/cli-installation-guide.md (reported line 112)May include surrounding context.

Option 2: Direct password (avoid on shared shells)

cloudrobo r2c client --bundle ./cert.zip --private-key-password

Option 3: Disable prompting (fail if encrypted)

cloudrobo r2c client --bundle ./cert.zip --no-prompt-password

text

The cloud observation workflow lacks a privacy warning for camera images and sensor data, including during dry-run testing.

Medium
Category
Output Handling
Confidence
92% confidence
Finding

The guide instructs users to start the client and verify cloud observation publishing, with earlier examples configuring camera capture and image mappings (L347–L349 and L426–L432). It explains that dry-run suppresses action execution but does not warn that observations, potentially including sensitive images of people or surroundings, are still transmitted to the configured cloud service.

Content

No source excerpt is available for this finding.