Back to skill

Security audit

free-llm-installer

Security checks for vulnerabilities and agentic risk

Overview

The skill is coherent, but it can persistently reroute Claude Code through local proxies and external AI services, including prompts and images, so it needs careful review before installation.

Install only if you are comfortable with this skill changing WorkBuddy and Claude Code model configuration, running a local proxy, and sending prompts or image content to third-party LLM providers. Avoid using it with secrets, private customer data, production code, or sensitive screenshots unless you have reviewed the target provider and disabled unneeded proxy, image-forwarding, watch, and failure-dump behavior.

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
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
  • System Prompt LeakageDirect Leakage, Indirect Extraction, Tool-Based Exfiltration
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • MCP Least PrivilegeUnderdeclared Capability, Wildcard Permission, Missing Permission Declaration
Findings (47)

Agent Config Directory Access

High
Category
Agent Snooping
Confidence
90% confidence
Finding

Skill reads from agent configuration directories (.claude/, .codex/, .gemini/). These directories may contain API keys, personal settings, and other credentials that the skill has no legitimate need to access.

Content

Scanner excerpt · README.md (reported line 61)May include surrounding context.

md
## 桌面切换器(tools/)

双击 `Claude模型切换器.cmd` → 弹出窗口,点哪个模型就切到哪个(写入
`~/.claude/settings.json` 并自动拉起需要的本地代理)→ **重启 Claude Code** 生效。

切换器支持:Kilo 文本 / 读图 Vision / DeepSeek-V4(付费,慎开自动模式) /
GLM-4.7 / GLM-4-Flash。要新增模型,在 `claude_model_switcher.pyw` 的

Tp4

High
Category
MCP Tool Poisoning
Confidence
96% confidence
Finding

The skill advertises free-model installation and verification, yet the documented behaviors include reading local images and sending them to third-party vision backends, plus optional reliance on API-keyed providers. Hidden data egress and undisclosed paid/keyed dependencies undermine informed consent and can expose sensitive local content under the guise of simple model setup.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding

The skill advertises free-model installation and verification, yet the documented behaviors include reading local images and sending them to third-party vision backends, plus optional reliance on API-keyed providers. Hidden data egress and undisclosed paid/keyed dependencies undermine informed consent and can expose sensitive local content under the guise of simple model setup.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
97% confidence
Finding

The skill advertises free-model installation and verification, yet the documented behaviors include reading local images and sending them to third-party vision backends, plus optional reliance on API-keyed providers. Hidden data egress and undisclosed paid/keyed dependencies undermine informed consent and can expose sensitive local content under the guise of simple model setup.

Content

No source excerpt is available for this finding.

External Script Fetching

High
Category
Supply Chain
Confidence
90% confidence
Finding

Remote code is downloaded and executed. This bypasses code review and could introduce malicious code.

Content

Scanner excerpt · SKILL.md (reported line 58)May include surrounding context.

md
> **正确做法:锁定具体模型 id**(如上面的 LFM2.5),别用 auto 路由。
> 探网关有什么可选(**免 key**):
> ```bash
> curl -s -H "Authorization: Bearer anonymous" \
>   https://api.kilo.ai/api/gateway/v1/models | python -c "import json,sys; [print(m['id'],m.get('context_length')) for m in json.load(sys.stdin)['data'] if 'free' in m['id']]"
> ```

Agent Config Directory Access

High
Category
Agent Snooping
Confidence
98% confidence
Finding

The skill instructs modification of ~/.claude/settings.json, which is a high-sensitivity agent configuration file controlling model routing, MCP servers, and behavior. Unauthorized or poorly scoped edits to this file can persistently alter agent trust boundaries, add new executable integrations, or reroute future conversations and tool use through attacker-controlled components.

Content

Scanner excerpt · SKILL.md (reported line 183)May include surrounding context.

1) (可选但推荐)设置智谱 key;不设置也能用,只是直接走 Kilo 兜底

export ZHIPU_API_KEY=你的智谱key # Windows: set ZHIPU_API_KEY=...

2) 在 ~/.claude/settings.json(或某个 profile)注入 mcpServers:

text
```json
{

Direct Prompt Extraction

High
Category
System Prompt Leakage
Confidence
85% confidence
Finding

Skill contains instructions that could directly expose system prompts, internal rules, or hidden instructions to users or external parties.

Content

Scanner excerpt · tools/ccproxy.py (reported line 1278)May include surrounding context.

python
oa_req = anthropic_to_openai(req)
        except Exception as e:
            self._send_error_anthropic(400, "invalid_request_error", "translate: %s" % e)
            return

        prompt_chars = sum(len(str(m.get("content", ""))) for m in oa_req["messages"])
        model_for_reply = req.get("model") or CONFIG["upstream"].get("model")

Agent Config Directory Access

High
Category
Agent Snooping
Confidence
95% confidence
Finding

The tool directly overwrites ~/.claude/settings.json and injects profile-controlled env, modelSettings, and mcpServers values. In the context of an agent skill, modifying the agent's config directory is high risk because it can silently reroute future model traffic, enable additional MCP servers, or persist altered behavior beyond the current run.

Content

Scanner excerpt · tools/claude_model_switcher.py (reported line 3)May include surrounding context.

python
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""Claude 模型切换器 —— 点哪个模型,就把 ~/.claude/settings.json 切到那个 profile,
并自动拉起对应的本地代理。双击「Claude模型切换器.cmd」启动。

本文件为可移植版本:python 取运行本脚本的解释器(pythonw),

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
98% confidence
Finding

The skill documentation is entirely written in Chinese, including setup and safety guidance, with no indication that users may choose another language or that the skill is intended only for a Chinese-speaking or region-specific audience. This creates a natural-language locale policy issue because it effectively imposes a language requirement without opt-in or justification.

Content

No source excerpt is available for this finding.

Undeclared Tool Scope

Medium
Category
MCP Least Privilege
Confidence
95% confidence
Finding

The skill describes capabilities that include shell commands, network access, reading/writing config files, and environment-variable use, but it declares no explicit tool scope or permission boundaries. That creates an overprivileged execution surface where an agent could perform sensitive local or outbound actions without a narrowly declared contract, increasing the chance of unsafe invocation or misuse.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

Grouping installation, verification, monitoring, and probing under one ambiguous invocation condition mixes passive and active behaviors. In practice, that means a user request that sounds informational could trigger invasive operations such as file edits, process launches, or network checks without a clear transition to consent.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
90% confidence
Finding

Grouping installation, verification, monitoring, and probing under one ambiguous invocation condition mixes passive and active behaviors. In practice, that means a user request that sounds informational could trigger invasive operations such as file edits, process launches, or network checks without a clear transition to consent.

Content

No source excerpt is available for this finding.

Ssd 3

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

The skill instructs persistent logging of model checks and status changes into files under the user's home directory. Those logs can accumulate prompts, endpoint details, operational metadata, failure traces, and other sensitive artifacts that may later be exposed to other local users, backup systems, or unrelated tools.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The manifest describes a skill for one-click installation, verification, and monitoring of free LLMs. However, the documentation and command examples also cover installing StepFun via --stepfun-key and even DeepSeek via --deepseek-key, with DeepSeek explicitly labeled paid, which goes beyond the stated free-LLM-focused purpose.

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 · SKILL.md (reported line 59)May include surrounding context.

md
# 内置配置(全部经实测验证)
# ---------------------------------------------------------------------------

KILO_URL = "https://api.kilo.ai/api/gateway/v1/chat/completions"
ZEN_URL = "https://opencode.ai/zen/v1/chat/completions"
ZHIPU_URL = "https://open.bigmodel.cn/api/paas/v4/chat/completions"
STEPFUN_URL = "https://api.stepfun.com/v1/chat/completions"

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 · SKILL.md (reported line 327)May include surrounding context.

md
# 内置配置(全部经实测验证)
# ---------------------------------------------------------------------------

KILO_URL = "https://api.kilo.ai/api/gateway/v1/chat/completions"
ZEN_URL = "https://opencode.ai/zen/v1/chat/completions"
ZHIPU_URL = "https://open.bigmodel.cn/api/paas/v4/chat/completions"
STEPFUN_URL = "https://api.stepfun.com/v1/chat/completions"

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 · SKILL.md (reported line 350)May include surrounding context.

md
# 内置配置(全部经实测验证)
# ---------------------------------------------------------------------------

KILO_URL = "https://api.kilo.ai/api/gateway/v1/chat/completions"
ZEN_URL = "https://opencode.ai/zen/v1/chat/completions"
ZHIPU_URL = "https://open.bigmodel.cn/api/paas/v4/chat/completions"
STEPFUN_URL = "https://api.stepfun.com/v1/chat/completions"

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 · examples/models.json.example (reported line 27)May include surrounding context.

text
# 内置配置(全部经实测验证)
# ---------------------------------------------------------------------------

KILO_URL = "https://api.kilo.ai/api/gateway/v1/chat/completions"
ZEN_URL = "https://opencode.ai/zen/v1/chat/completions"
ZHIPU_URL = "https://open.bigmodel.cn/api/paas/v4/chat/completions"
STEPFUN_URL = "https://api.stepfun.com/v1/chat/completions"

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 · examples/models.json.example (reported line 48)May include surrounding context.

text
# 内置配置(全部经实测验证)
# ---------------------------------------------------------------------------

KILO_URL = "https://api.kilo.ai/api/gateway/v1/chat/completions"
ZEN_URL = "https://opencode.ai/zen/v1/chat/completions"
ZHIPU_URL = "https://open.bigmodel.cn/api/paas/v4/chat/completions"
STEPFUN_URL = "https://api.stepfun.com/v1/chat/completions"

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 · scripts/install_free_llms.py (reported line 78)May include surrounding context.

python
# 内置配置(全部经实测验证)
# ---------------------------------------------------------------------------

KILO_URL = "https://api.kilo.ai/api/gateway/v1/chat/completions"
ZEN_URL = "https://opencode.ai/zen/v1/chat/completions"
ZHIPU_URL = "https://open.bigmodel.cn/api/paas/v4/chat/completions"
STEPFUN_URL = "https://api.stepfun.com/v1/chat/completions"

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 · SKILL.md (reported line 251)May include surrounding context.

md
KILO_URL = "https://api.kilo.ai/api/gateway/v1/chat/completions"
ZEN_URL = "https://opencode.ai/zen/v1/chat/completions"
ZHIPU_URL = "https://open.bigmodel.cn/api/paas/v4/chat/completions"
STEPFUN_URL = "https://api.stepfun.com/v1/chat/completions"

# Kilo / Zen 前面有 Cloudflare,按 User-Agent 拦截,必须带浏览器 UA
UA = ("Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 "

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 · scripts/install_free_llms.py (reported line 81)May include surrounding context.

python
KILO_URL = "https://api.kilo.ai/api/gateway/v1/chat/completions"
ZEN_URL = "https://opencode.ai/zen/v1/chat/completions"
ZHIPU_URL = "https://open.bigmodel.cn/api/paas/v4/chat/completions"
STEPFUN_URL = "https://api.stepfun.com/v1/chat/completions"

# Kilo / Zen 前面有 Cloudflare,按 User-Agent 拦截,必须带浏览器 UA
UA = ("Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 "

Description-Behavior Mismatch

Medium
Category
Not specified by scanner
Confidence
88% confidence
Finding

Beyond installing and validating configured free models, the code probes multiple third-party platforms for reachability and enumerates OpenRouter's free model catalog. While some monitoring is described in the manifest, this broader external platform discovery capability is more of a network reconnaissance/reporting feature than an obvious implementation detail of installing local WorkBuddy/Claude Code models.

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 · scripts/install_free_llms.py (reported line 429)May include surrounding context.

python
FOREIGN_PLATFORMS = [
    ("OpenRouter", "https://openrouter.ai/api/v1/models", False,
     "国内可达,公开列表免 key;调用仍需 OpenRouter Key"),
    ("Groq", "https://api.groq.com/openai/v1/models", True,
     "需 key,国内常被 CF 拦"),
    ("Together", "https://api.together.xyz/v1/models", True,
     "需 key,国内常被 CF 拦"),

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 · scripts/install_free_llms.py (reported line 431)May include surrounding context.

python
"国内可达,公开列表免 key;调用仍需 OpenRouter Key"),
    ("Groq", "https://api.groq.com/openai/v1/models", True,
     "需 key,国内常被 CF 拦"),
    ("Together", "https://api.together.xyz/v1/models", True,
     "需 key,国内常被 CF 拦"),
    ("HuggingFace", "https://router.huggingface.co/api/v1/models", True,
     "需 HF Token,国内常被代理 502"),