Back to skill

Security audit

arkcli-train-finetune

Security checks for vulnerabilities and agentic risk

Overview

ClawScan marks this skill suspicious because it reports an unpinned SDK download, an authentication command that conflicts with its own guidance, undisclosed activity reporting, and a pipeline testing command outside the stated fine-tuning purpose.

The verdict is suspicious, so consider avoiding installation until the SDK source is pinned and its integrity can be verified, the conflicting authentication example is resolved, and activity reporting and remote data submission are clearly disclosed. The scanner reports full component coverage, but Skillspector also returned an exit-status error with serializer warnings, so treat the assessment as informative rather than proof that every risk has been ruled out.

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
  • YARA SignaturesMalware Match, Webshell Match, Cryptominer Match
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
Findings (7)

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
89% confidence
Finding

The supplied command concerns pipeline-plugin testing rather than fine-tuning jobs or their associated data, metrics, artifacts, and deployments. Although help queries can support a declared workflow, this chunk shows no connection between that testing command and the stated purpose.

Content

No source excerpt is available for this finding.

YARA rule 'agent_skill_remote_bootstrap_execution': Remote script or code download followed by execution/bootstrap installation [agent_skills]

High
Category
YARA Match
Confidence
96% confidence
Finding

The installation command retrieves a mutable latest source archive without a pinned version or independently verified integrity value; installing a source distribution can execute build code. Compromise of the distribution location could therefore execute code with the agent user's privileges and access local credentials; a virtual environment and installation consent do not provide a security sandbox or verify the artifact.

Content

Scanner excerpt · references/ark-finetune-sdk.md (reported line 31)May include surrounding context.

��较快的模型支持列表或敏感来源链接写进用户可见输出或生成文件。

起步检查

处理任何实质性任务前:

  1. 准备 venv 并安装 SDK(用户自管 venv,arkcli 不代管)

    bash
    python3.11 -m venv /tmp/ark-rl     # 必须 ≥ 3.10
    source /tmp/ark-rl/bin/activate
    

    火山方舟 SDK 使用以下安装源:

    bash
    python -m pip install https://ark-public-example-cn-beijing.tos-cn-beijing.volces.com/ark-sdk/ark_sdk-latest.tar.gz
    

    后续 ark / python 命令必须在同一 shell 里执行,或显式用 /tmp/ark-rl/bin/ark。

  2. 完成 SDK 鉴权 火山方舟鉴权走 ArkCLI 桥,不要单独执行 ark login:

    bash
    arkcli auth login                            # 已登录时跳过
    arkcli train finetune sdk login              # 把 ArkCLI STS 写入 ~/.ark/authorization.json
    

    说明:

    • sdk login 会临时覆盖 ~/.ark/authorization.json;如果用户机器上

The common-command shell snippet includes `ark login`, contradicting the explicit prohibition against using it.

Medium
Category
Excessive Agency
Confidence
93% confidence
Finding

The common-command block includes an authentication command that the same document explicitly prohibits because it overwrites the bridged STS authentication state. An agent copying that block could disrupt authentication or create account-context confusion, although the earlier warnings reduce the likelihood and no credential disclosure is demonstrated.

Content

No source excerpt is available for this finding.

Workflow activity reporting is required without a user-facing telemetry disclosure.

Medium
Category
Output Handling
Confidence
90% confidence
Finding

The workflow mandates activity-reporting calls with a persistent flow identifier without explaining that reporting to the user. This introduces an undisclosed telemetry side effect that can reveal workflow entry and validation outcomes, but the supplied text does not show transmission of training contents or credentials.

Content

No source excerpt is available for this finding.

The submission workflow lacks a warning about sending local training data to the remote training service.

Medium
Category
Output Handling
Confidence
90% confidence
Finding

The workflow instructs the agent to submit job.yaml, whose example training configuration references ./data/train.jsonl (L084–L087). It does not warn that submission can upload local training data for remote processing, which matters if the dataset contains confidential or personal information.

Content

No source excerpt is available for this finding.

The troubleshooting workflow requires activity reporting unrelated to diagnosing or managing training jobs.

Medium
Category
Excessive Agency
Confidence
93% confidence
Finding

The troubleshooting procedure requires a separate activity-reporting operation once a diagnostic conclusion is reached, even though that operation is not necessary to diagnose the job. This adds an undisclosed side effect to an otherwise observational workflow; the visible information is a debug-success event and flow linkage, not logs or credentials.

Content

No source excerpt is available for this finding.

The troubleshooting workflow requires activity reporting without a user-facing telemetry disclosure.

Medium
Category
Output Handling
Confidence
91% confidence
Finding

The required debug_success reporting call has no accompanying user-facing telemetry disclosure and reuses or creates a flow identifier. It can expose troubleshooting activity and correlate workflow events; this is the same reporting operation identified by the companion finding, not evidence of an additional data-transfer channel.

Content

No source excerpt is available for this finding.