Back to skill

Security audit

cargo-orchestration

Security checks for vulnerabilities and agentic risk

Overview

Review recommended: conflicting record-selection instructions and incomplete replay safeguards could cause unintended processing or repeated business actions.

Before using this with live business data, replace empty filters with the intended audience and approve the exact records and cost. Check completed steps before replaying failures so messages or CRM changes are not repeated. Use only approved files and provider connections, restrict account permissions, and pin a reviewed CLI version. Material scanner gaps were checked directly, but the external CLI and missing sibling prerequisite guidance remain unverified.

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
  • 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 (36)

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

The supplied commands manage segmentation resources rather than execute or inspect workflows and runs. Segment updates and deletion are distinct capabilities not disclosed in the declared purpose; using the same Cargo CLI does not make them supporting implementation details of the stated functions.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding

The clearest mismatch is the inclusion of record download and download-outputs commands despite the description explicitly excluding result-file downloads. The chunk also exposes record cancellation, which goes beyond the described execution, construction, visualization, and history-query capabilities. Listing, filtering, counting, and inspecting execution metrics are reasonably related to the declared run-history functionality and are not independently flagged.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
96% confidence
Finding

Most commands align with the declared orchestration purpose or support it. However, deleting plays and tools is a distinct destructive resource-management capability not disclosed by the description. Building or editing workflows does not adequately describe deleting these resources. Play disabling and health-alert configuration also extend into administrative management, but the explicit deletion commands are the clearest mismatch.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
95% confidence
Finding

The supplied chunk performs a standalone data-retrieval operation, not any of the declared execution, workflow-management, visualization, agent-messaging, or run-history operations. Fetching records could support a batch workflow, but no such workflow or execution step appears in this chunk.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
95% confidence
Finding

The supplied command cancels workflow runs rather than executing, building, visualizing, or querying them. Cancellation is a distinct state-changing capability that is not disclosed in the declared description.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
95% confidence
Finding

The supplied commands use cargo-ai segmentation segment fetch to retrieve records. They do not execute actions or workflows, launch batches, interact with agents, manipulate workflow definitions, or query run history. Although fetching records could support a batch workflow, this chunk presents standalone segment-retrieval examples without connecting them to a declared capability.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
95% confidence
Finding

The supplied command stops an existing batch rather than launching or executing one. Cancellation is a distinct state-changing operation, not merely an implementation detail of the declared execution, editing, visualization, or querying capabilities.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
94% confidence
Finding

Cancellation is a distinct state-changing capability not disclosed in the description. These commands stop existing executions rather than launch, build, visualize, or query them, so they are not merely supporting implementation details of the declared functions.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
88% confidence
Finding

The supplied commands perform segmentation discovery and record retrieval, rather than executing, building, visualizing, or inspecting the history of a workflow. Although retrieving records could support a batch operation, this chunk contains no such operation or explicit connection to one. Its demonstrated capability is not covered by the declared purpose.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
85% confidence
Finding

Enabling or disabling plays and configuring health alerts are reasonably related to the declared orchestration and scheduling functions. Permanently removing a play, however, is a distinct destructive capability beyond the stated execution, editing, visualization and querying functions. The supplied chunk contains command examples rather than evidence that these commands have actually been executed.

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 99)May include surrounding context.

md
> `references/examples/templates.md` — pre-built workflow templates

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 148)May include surrounding context.

md
suited for tools (`"kind":"tool"`) or plays (`"kind":"play"`) right away. See `references/examples/templates.md` for the full guide.

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 550)May include surrounding context.

md
See `references/examples/templates.md` for the full guide including placeholder conventions and end-to-end examples.

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 101)May include surrounding context.

md
> `references/examples/segments.md` — segment fetch and filter examples

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 528)May include surrounding context.

md
and `--sync`. See `references/filter-syntax.md` for the full filter syntax and `references/examples/segments.md` for filtering, pagination, sorting, enrollment

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

Medium
Category
MCP Rug Pull
Confidence
93% confidence
Finding

The documented npx invocation can download and execute an unpinned CLI release, making the executed code dependent on mutable registry state. A compromised release could execute with the agent's local privileges and access available credentials; this is a CLI supply-chain issue, not specifically an MCP-server issue.

Content

No source excerpt is available for this finding.

The mandatory sample workflow sends live records to external services without a privacy warning.

Medium
Category
Output Handling
Confidence
92% confidence
Finding

The workflow instructs the agent to process 10–20 real records and, for file sources, upload a CSV containing lead data to Cargo. It warns about sampling costs and requires approval for full enrollment, but does not disclose that sample records leave the user's environment and may also reach connector providers used by the workflow.

Content

No source excerpt is available for this finding.

The skill provides result-file download and bulk export instructions despite explicitly routing result downloads to cargo-analytics.

Medium
Category
Excessive Agency
Confidence
98% confidence
Finding

The manifest at L003 explicitly says to skip this skill when downloading result files and use cargo-analytics instead. Nevertheless, L489–L496 provide executable record-download and paginated output-export commands, with additional download instructions at L442 and L472, implementing the excluded task rather than handing it off.

Content

No source excerpt is available for this finding.

Deleting standalone segments is outside the skill's stated workflow-orchestration purpose.

Medium
Category
Excessive Agency
Confidence
91% confidence
Finding

The manifest covers executing actions, building or editing workflows, scheduling automation, messaging agents, visualization, and run-history queries—not deleting segmentation resources. The segment removal command deletes a standalone resource and explicitly cannot operate on a segment linked to a workflow, making it distinct from maintaining the workflow being orchestrated.

Content

No source excerpt is available for this finding.

The lead-enrichment workflow submits personal data without a privacy warning.

Medium
Category
Output Handling
Confidence
91% confidence
Finding

The end-to-end workflow instructs execution of a hosted enrichment tool with leads' email addresses and company names. It does not warn that using real records transmits personal data outside the local environment or explain that the selected tool may forward it to enrichment providers.

Content

No source excerpt is available for this finding.

The action-enabled examples instruct the agent to enrich lead data and modify Salesforce records without a safety warning.

Medium
Category
Output Handling
Confidence
91% confidence
Finding

This example supplies executable enrichment and Salesforce actions and explicitly requests adding a lead, rather than merely listing available tools. Neither this workflow nor the combined CRM-update example warns that execution can share lead data with connected services and change live CRM records.

Content

No source excerpt is available for this finding.

The knowledge-file upload workflow lacks a warning about transferring potentially confidential documents.

Medium
Category
Output Handling
Confidence
94% confidence
Finding

The workflow instructs uploading a local PDF and suggests internal ICP criteria or product information as inputs. It does not warn that the file leaves the local system or advise checking whether its contents are approved for sharing with the service.

Content

No source excerpt is available for this finding.

The lead-scoring workflow sends model data to OpenAI without a privacy warning.

Medium
Category
Output Handling
Confidence
96% confidence
Finding

The prompt includes company names, domains, and employee counts and configures an OpenAI connector to process them. Steps 6–7 instruct running these nodes against every model row, but the file does not warn that these fields will be transmitted to an external AI provider.

Content

No source excerpt is available for this finding.

The end-to-end example claims to run against the play's segment but executes against every row in the model.

Medium
Category
Excessive Agency
Confidence
99% confidence
Finding

The end-to-end description promises execution against the play's segment, but the commands use an empty filter that selects every row in the model. An agent following that description could process records outside the authorized audience, incurring unintended model costs and sending additional records to the configured AI provider; inline comments acknowledging all rows do not resolve the conflicting scope.

Content

No source excerpt is available for this finding.

The template execution workflow lacks a privacy warning that input data is sent to Clearbit.

Medium
Category
Output Handling
Confidence
92% confidence
Finding

The workflow instructs the agent to execute the company-enrichment template with a supplied domain; the connector configuration at L079–L087 passes that domain to Clearbit. Although Clearbit is named in the template description, the execution instructions do not warn that user-supplied input will be transmitted to a third-party service.

Content

No source excerpt is available for this finding.