Back to skill

Security audit

WorkBuddy 自定义模型能力位排查(读图/工具/推理)

Security checks for vulnerabilities and agentic risk

Overview

ClawScan rates this skill suspicious because scanners report broad cleanup that may delete unrelated temporary backups and configuration changes that may discard unrelated settings; analysis was partial.

Consider avoiding installation until the reported cleanup and configuration-write behavior has been reviewed or corrected. If you use it, check what it may change or remove, avoid sensitive images in model tests unless you understand where they are sent, and provide only the access needed. The skill scan covered 75% of components; it partially inspected SKILL.md, did not analyze the bundled image, and could not verify the imported probe_model implementation's additional behavior. The separate AIG scan completed with no findings, but that does not resolve these gaps.

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 Least PrivilegeUnderdeclared Capability, Wildcard Permission, Missing Permission Declaration
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
Findings (9)

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
97% confidence
Finding

The capability-configuration tests support the declared purpose and are not themselves a mismatch. However, bak_dir is constructed from the parent of a tempfile.mkstemp file, typically producing /tmp/backups rather than an isolated test directory. The final cleanup enumerates that directory and calls os.unlink on every entry before removing the directory. This can delete unrelated backup files and exceeds cleanup of test-owned artifacts. The imported probe_model implementation is not supplied, so its additional behavior cannot be verified.

Content

No source excerpt is available for this finding.

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

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.

External Transmission

Medium
Category
Data Exfiltration
Confidence
50% confidence
Finding

Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.

Content

Scanner excerpt · SKILL.md (reported line 64)May include surrounding context.

md
# 加一个全新模型:先追加条目,再探测,再按探测结果填能力位
python probe_model.py --add --model my-new-model \
  --url https://api.example.com/chat/completions --key sk-xxx \
  --name "My Model" --vendor "Example"

# 常用组合

The image-testing workflow uploads a local image to the configured model endpoint without a privacy warning.

Medium
Category
Output Handling
Confidence
97% confidence
Finding

The direct-test example reads a local image and transmits its complete contents to a model provider without warning against using sensitive images. Although remote image submission is relevant to capability testing, an agent following this workflow could disclose private attachments unnecessarily; the generated synthetic test image is a safer alternative.

Content

No source excerpt is available for this finding.

The --add operation persists an untested model before probing, contrary to its documented probe-before-write behavior.

Medium
Category
Excessive Agency
Confidence
99% confidence
Finding

The --add path calls append_entry before making probe requests, so it persists the endpoint and credentials even if probing fails or apply_results subsequently refuses to write. This violates the advertised probe-before-write behavior and leaves a live configuration change outside the probe-success safeguards, although adding the model is user-requested.

Content

No source excerpt is available for this finding.

Configuration normalization deletes unrelated fields despite the promise to change only capability settings.

Medium
Category
Excessive Agency
Confidence
99% confidence
Finding

Both write paths reconstruct every model entry using only FIELD_ORDER, silently dropping all unlisted fields, including those on models not being probed. This exceeds the promised capability-only modification and can destroy unrelated configuration or remove settings with security significance; backups mitigate recovery costs but do not prevent the damage.

Content

No source excerpt is available for this finding.

Test cleanup deletes unrelated files from a shared temporary backups directory (L177–L179).

Medium
Category
Excessive Agency
Confidence
100% confidence
Finding

The temporary model files are placed directly in the system temporary directory, making bak_dir a shared backups subdirectory rather than a test-owned directory. Cleanup unlinks every file there without checking ownership, potentially deleting unrelated backups accessible to the executing user.

Content

No source excerpt is available for this finding.

Cleanup deletes every file in a shared temporary backups directory without disclosure.

Medium
Category
Output Handling
Confidence
100% confidence
Finding

Running the regression test triggers deletion of every file in the shared temporary backups directory, an unexpected destructive side effect unrelated to the test's own artifacts. This is the same underlying cleanup defect as the other finding at these lines, not an additional independent deletion mechanism.

Content

No source excerpt is available for this finding.

External Transmission

Medium
Category
Data Exfiltration
Confidence
60% confidence
Finding

Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.

Content

Scanner excerpt · test_vision_stability.py (reported line 17)May include surrounding context.

python
import probe_model as pm

MODEL = os.environ.get("PROBE_MODEL", "deepseek-v4-flash")
URL = os.environ.get("PROBE_URL", "https://api.deepseek.com/chat/completions")
# ⚠ key 只能从环境变量读。此前这个文件里硬编码过一枚真实 key,
# skill 一旦发布就等于公开泄露 —— 已改为必须显式提供。
KEY = os.environ.get("DEEPSEEK_API_KEY") or os.environ.get("PROBE_KEY")