Back to skill

Security audit

agent-mailbox

Security checks for vulnerabilities and agentic risk

Overview

ClawScan rates this skill suspicious because scanners report risks around bypassing email-send confirmation, running unverified installers, and changing npm settings; the review is incomplete because some referenced local paths were not found.

Review the email confirmation and recipient rules, the downloaded installer, and the persistent npm registry change before installing. Consider whether you are comfortable with the reported ability to send email without a separate review, and with setup guidance that includes sudo. The scan was partial: three local path references could not be resolved, so the evidence does not cover those referenced paths.

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
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
  • Output HandlingUnvalidated Output Injection, Cross-Context Output, Unbounded Output
  • Tool MisuseTool Parameter Abuse, Chaining Abuse, Unsafe Defaults
Findings (14)

Chaining Abuse

High
Category
Tool Misuse
Confidence
70% confidence
Finding

Tool calls are chained to bypass individual safety checks or escalate capabilities beyond what any single tool call would allow.

Content

Scanner excerpt · README.md (reported line 33)May include surrounding context.

bash
# Linux / macOS
curl -o- https://raw.githubusercontent.com/nvm-sh/nvm/v0.40.1/install.sh | bash
source ~/.bashrc
nvm install 20
node -v

Chaining Abuse

High
Category
Tool Misuse
Confidence
70% confidence
Finding

Tool calls are chained to bypass individual safety checks or escalate capabilities beyond what any single tool call would allow.

Content

Scanner excerpt · SKILL.md (reported line 32)May include surrounding context.

没有 Node(或报 exec: node: not found):

bash
curl -o- https://raw.githubusercontent.com/nvm-sh/nvm/v0.40.1/install.sh | bash
source ~/.bashrc
nvm install 20

Dependency Source Redirection

High
Category
Supply Chain
Confidence
100% confidence
Finding

Dependency resolution is redirected away from a canonical default, adds another source, or uses a destination that cannot be resolved statically.

Content

Scanner excerpt · README.md (reported line 51)May include surrounding context.

md
npm config set registry https://registry.npmmirror.com

A troubleshooting pretext pressures the agent to bypass or immediately complete the email confirmation step without review.

High
Category
Prompt Injection
Confidence
96% confidence
Finding

The same-shell requirement and prohibition on intervening log checks pressure the agent to complete sending immediately or bypass confirmation entirely. The stated token lifetime does not itself establish those restrictions, and this workflow discourages review before a potentially irreversible disclosure.

Content

No source excerpt is available for this finding.

The Node setup instructions execute a downloaded script without warning about remote code execution or local configuration changes.

Medium
Category
Output Handling
Confidence
94% confidence
Finding

The installation command executes downloaded shell code with the current user's permissions without an independent integrity check or review. Although the URL points to a recognizable upstream project over HTTPS, compromised upstream content could access local credentials and modify shell configuration.

Content

No source excerpt is available for this finding.

The documented send command bypasses the two-step confirmation promised by the manifest.

Medium
Category
Excessive Agency
Confidence
97% confidence
Finding

The documentation recommends --confirmed as a shortcut without requiring prior approval of the actual message. In an agent-operated mailbox, this can remove an opportunity to catch unintended recipients, sensitive attachments, or unauthorized sends; a dry run alone does not establish user consent.

Content

No source excerpt is available for this finding.

Sudo/Root Execution

Medium
Category
Privilege Escalation
Confidence
87% confidence
Finding

The troubleshooting table offers privileged npm installation as a routine response to a permissions error. This unnecessarily grants the installation process elevated access and increases the consequences of a compromised package or installer; the nonprivileged alternative reduces but does not eliminate that risk.

Content

Scanner excerpt · README.md (reported line 137)May include surrounding context.

md
|---|---|---|
| `exec: node: not found` | 有 npm 没 node / PATH 未刷新 | 装 Node 后**重开终端**;nvm 用户先 `source ~/.nvm/nvm.sh` |
| `npm install` 卡住 | 官方源慢 | `npm config set registry https://registry.npmmirror.com` |
| `EACCES: permission denied` | 全局目录无权限 | `sudo npm install -g ...`;或 `npm config set prefix ~/.npm-global` |
| `agently-cli: command not found` | npm 全局 bin 不在 PATH | `npm bin -g` 查看并加入 PATH |
| 授权链接打不开 | 链接一次性、已过期 | 重跑 `agently-cli auth login` |
| 授权成功但命令报 401 | 本地 token 失效 | `agently-cli auth refresh`,仍不行则 `auth logout` 后重新 login |

The mandatory CC rule lacks a privacy warning about disclosing email contents to additional recipients.

Medium
Category
Output Handling
Confidence
94% confidence
Finding

The blanket instruction to copy a supervisor or audit address expands the audience for every external message without checking whether that recipient is authorized for its contents. In a mailbox handling private correspondence and attachments, this can cause unintended disclosure, although no attacker-controlled destination is specified.

Content

No source excerpt is available for this finding.

The Node.js installation workflow executes a downloaded shell script without warning about its access to the local environment.

Medium
Category
Output Handling
Confidence
94% confidence
Finding

The agent is instructed to execute a downloaded installer without verifying it or explaining its access to the local environment. The script runs with the agent's user permissions, so compromised upstream content could access credentials and alter startup files.

Content

No source excerpt is available for this finding.

The registry-switch instruction omits disclosure that it persistently redirects npm package requests to a third-party mirror.

Medium
Category
Output Handling
Confidence
97% confidence
Finding

The network workaround silently makes a persistent user-level npm registry change to a third-party service. An agent following it can redirect unrelated future installs through that service, extending the supply-chain trust boundary beyond the mailbox setup.

Content

No source excerpt is available for this finding.

The skill advertises two-step confirmation for sending and replying but explicitly instructs agents how to bypass it.

Medium
Category
Excessive Agency
Confidence
97% confidence
Finding

The skill explicitly tells agents to skip the advertised confirmation step with --confirmed without conditioning this on approval of the final message. Because sending mail is an external action, this can turn an agent mistake or an instruction from untrusted email content into an immediate unauthorized send.

Content

No source excerpt is available for this finding.

Sudo/Root Execution

Medium
Category
Privilege Escalation
Confidence
90% confidence
Finding

The skill directs an agent to resolve npm permission failures using sudo without an approval gate or a least-privilege alternative. Elevating the installation process increases the potential damage from compromised installation inputs or unintended system changes.

Content

Scanner excerpt · SKILL.md (reported line 106)May include surrounding context.

md
| 报错 | 原因 | 处理 |
|---|---|---|
| `exec: node: not found` | 缺 Node / PATH 未刷新 | 装 Node 后重开终端 / `source ~/.nvm/nvm.sh` |
| `EACCES` | 全局目录权限 | `sudo npm i -g ...` |
| `command not found` | npm bin 不在 PATH | `npm bin -g` 加入 PATH |
| 401 / Invalid token | 授权失效 | `auth refresh` → 仍失败则 `auth logout` + `auth login` |
| 链接打不开 | 授权链接一次性且已过期 | 重跑 `auth login` |

External Script Fetching

Low
Category
Supply Chain
Confidence
94% confidence
Finding

The command pipes an external installer directly into bash, allowing the returned content to execute before inspection. The recognizable HTTPS upstream reduces some risks but does not independently verify the script's integrity or prevent upstream compromise.

Content

Scanner excerpt · README.md (reported line 33)May include surrounding context.

bash
# Linux / macOS
curl -o- https://raw.githubusercontent.com/nvm-sh/nvm/v0.40.1/install.sh | bash
source ~/.bashrc
nvm install 20
node -v

External Script Fetching

Low
Category
Supply Chain
Confidence
95% confidence
Finding

The agent setup command immediately executes whatever shell content the remote installer URL returns. HTTPS and a versioned upstream path help establish transport and source expectations, but neither provides independent artifact verification.

Content

Scanner excerpt · SKILL.md (reported line 32)May include surrounding context.

没有 Node(或报 exec: node: not found):

bash
curl -o- https://raw.githubusercontent.com/nvm-sh/nvm/v0.40.1/install.sh | bash
source ~/.bashrc
nvm install 20