Back to skill

Security audit

cargo-segmentation

Security checks for vulnerabilities and agentic risk

Overview

The skill supports normal Cargo audience management, but instructs the agent to submit potentially sensitive diagnostics without user review.

Before installing, require sanitized support reports to be shown for approval. Explicitly authorize full exports and segment changes, use restricted Cargo permissions, and pin a reviewed CLI version. The external CLI and omitted sibling instructions were not audited.

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
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
Findings (8)

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

Support reporting is a legitimate ancillary operation, but the skill automatically directs submission of commands, verbatim errors, and workspace identifiers after repeated failures without redaction or user review. Commands can contain API tokens or sensitive filter values, so this can disclose secrets or customer information to a remote reporting system.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding

Retrieving segment metadata and parsing its filter are consistent supporting operations, but the final download command performs an export-like operation rather than segment creation, counting, or membership-change inspection. This crosses the description's explicit boundary assigning row exports to cargo-analytics.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
97% confidence
Finding

Most commands support the declared segmentation purpose. However, the download command exposes a row-export capability that the description specifically directs to another skill. This is a substantive scope mismatch rather than a supporting implementation detail.

Content

No source excerpt is available for this finding.

MCP server referenced without pinned version: 'npx @cargo-ai/cli'.

Medium
Category
MCP Rug Pull
Confidence
90% confidence
Finding

The bootstrap instructions install or execute the CLI without a fixed version, and the installation metadata explicitly selects latest. This creates a supply-chain exposure: a compromised future release could execute code with the agent's local privileges and access available credentials; this is an npm dependency issue, not evidence of an MCP server.

Content

No source excerpt is available for this finding.

The skill directly instructs full-audience exports despite explicitly excluding exports from its manifest scope.

Medium
Category
Excessive Agency
Confidence
98% confidence
Finding

The manifest at L003 says exporting segment rows belongs to cargo-analytics. However, L135 describes obtaining signed URLs for the full dataset, and L140 directly instructs using download for records beyond a preview. This goes beyond inspecting or counting segment membership and implements the explicitly excluded export workflow.

Content

No source excerpt is available for this finding.

Repeated CLI failures trigger remote support-report creation containing command and workspace details.

Medium
Category
Excessive Agency
Confidence
94% confidence
Finding

Two failed commands trigger remote report creation containing commands run, verbatim error messages, and UUIDs, with no sanitization or approval step. Those diagnostics may include token-bearing commands, sensitive query values, or returned customer data, allowing accidental disclosure beyond the segmentation task.

Content

No source excerpt is available for this finding.

The skill documents full-dataset exports despite explicitly assigning exports to cargo-analytics.

Medium
Category
Excessive Agency
Confidence
92% confidence
Finding

The manifest says to skip this skill when exporting segment rows and use cargo-analytics instead. This reference nevertheless exposes segmentation segment download and instructs the agent to fetch signed URLs containing the full dataset, providing an export workflow outside the stated segmentation scope.

Content

No source excerpt is available for this finding.

The skill provides an executable segment-download recipe despite explicitly excluding row exports from its scope.

Medium
Category
Excessive Agency
Confidence
98% confidence
Finding

The manifest explicitly directs exporting segment rows to cargo-analytics. This troubleshooting section instead instructs the agent to retrieve a saved segment's model and filter and invoke cargo-ai segmentation segment download, implementing the excluded export workflow rather than handing it off.

Content

No source excerpt is available for this finding.