Back to skill

Security audit

Milestone Execution

Security checks for vulnerabilities and agentic risk

Overview

The skill describes a powerful staged task executor, but its core executable is missing and the instructions would run an unverified command from the environment.

Install only after the package includes the reviewed `milestone-executor` implementation and uses a package-relative invocation path; also review the local state file behavior and require explicit confirmation for continuation, rollback, recovery, and any parallel work sessions.

Vulnerability Patterns
  • Tool Hijacking and SpoofingModifies or replaces tools so legitimate-looking calls execute attacker logic
  • 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
Findings (1)

T07 · Tool Hijacking and Spoofing

Error
Location
SKILL.md:31
Finding

Unbundled Core Executable Is Invoked Through PATH Resolution

Content
View full analysis

Vulnerability Details

File Location: SKILL.md:31-42, corroborated by README.md:132-137 and scripts/start.js:14-21
Vulnerability Type: Untrusted executable resolution through the system PATH
Risk Level: High

Vulnerable Code

SKILL.md:31-42 directs the agent to invoke the executable by its bare command name:

text
| User instruction | Main-session behavior |
|------------------|-----------------------|
| `Start [task description]` | Run `milestone-executor start "task" "m1|m2|m3"` |
| `Continue` | Run `milestone-executor continue <workSessionKey>` |
| `Stop` | Run `milestone-executor stop <workSessionKey>` |
| `Report` | Run `milestone-executor status` |
| `retry` | Run `milestone-executor retry <workSessionKey>` |
| `skip` | Run `milestone-executor skip <workSessionKey>` |
| `rollback` | Run `milestone-executor rollback` |
| `Modify [content]` | Run `milestone-executor modify <workSessionKey> "content"` |
| `Recover` | Run `milestone-executor recover` |

README.md:132-137 represents the executable as part of the package, but it is absent from the audited directory:

text
milestone-execution/
├── SKILL.md                    # Skill definition
├── README.md                   # This file
└── scripts/
    ├── milestone-executor      # Core execution script
    ├── start.js                # Deprecated launcher
    └── state.js                # Deprecated state manager

scripts/start.js:14-21 also redirects users to the missing executable:

javascript
console.log(`
⚠️ start.js is deprecated

Please use milestone-executor:
  milestone-executor start "task description" "m1|m2|m3"

Or use the full path:
  ~/.openclaw/skills/milestone-execution/scripts/milestone-executor start "task" "m1|m2|m3"
`);

Technical Analysis

The documented core executable, milestone-executor, is not present in the audited package ...[truncated 2532 chars]

Remediation
View remediation

Remediation Suggestions

  1. Bundle the reviewed milestone-executor implementation inside the package at the location declared by the documentation.
  2. Invoke it through a canonical package-relative or resolved absolute path rather than through PATH lookup, for example by resolving the path from the skill installation directory.
  3. Before execution, verify that the target exists, is a regular file, has expected ownership and permissions, and is not a symbolic link to an untrusted location.
  4. Add an integrity check using a trusted manifest or cryptographic digest distributed with the package.
  5. Fail closed if the bundled executable is missing or fails verification. Do not fall back to an executable found elsewhere in PATH.
  6. Remove or update documentation that lists files not included in the artifact.
  7. Use a restricted execution environment with a minimal trusted PATH and least-privilege filesystem, credential, network, and session permissions.
  8. Add package tests that verify all documented executable paths exist and resolve within the installed package.
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
Findings (10)

Tp4

High
Category
MCP Tool Poisoning
Confidence
97% confidence
Finding

The declared description promises a functioning staged task executor with milestone-based pauses and user-controlled progression. However, the supplied code chunk does not implement that behavior at all. It simply logs that start.js is deprecated, suggests an alternative command, and exits. This is a materially different primary purpose from the declared functionality, so it should be flagged as a mismatch.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
91% confidence
Finding

声明的核心价值是“多阶段任务执行控制器”,强调 AI 实际按阶段工作,并在人机确认点暂停继续;但代码仅提供 milestone 状态的 CLI 管理,包括 init/status/update/clear/rollback。它没有任何任务执行逻辑、没有用户确认流程、没有独立工作会话控制,也没有自动在 milestone 完成后暂停汇报的机制。虽然状态管理可能是该技能的支撑组件,但就所给代码块本身而言,其实际主功能更接近‘任务进度状态文件管理器’而非声明中的‘多阶段任务执行器’。因此存在明显描述与行为不一致。

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

The activation interface relies on very broad natural-language commands like “开始”, “继续”, and “修改 [内容]” without clear scoping, authorization, or confirmation boundaries. In a multi-turn agent environment, such generic triggers can be invoked accidentally or by prompt-injected content, causing unintended task execution, continuation, or workflow changes.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
87% confidence
Finding

Exposing rollback without a clear warning can cause destructive or confusing reversion of prior work, especially when combined with autonomous file modifications across milestones. In this skill's context, rollback affects task state and potentially user work products, so an operator may trigger data loss or inconsistent repository state without understanding the consequences.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The manifest description and earlier architecture emphasize that the AI executes milestones in a single independent work session, pausing after each milestone for user confirmation. However, the '并行 Milestone' section explicitly allows two milestones to run simultaneously in their own separate work sessions, which materially expands the execution model beyond the stated single-session staged flow.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

Lines L186-L190 declare a forced rule that all milestones execute in one work session, and L200 reinforces 'one task = one work session = multiple milestones.' In direct contradiction, the parallel execution section states that two milestones run simultaneously in their respective work sessions, so the documentation gives mutually incompatible implementation intent.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Low
Category
Not specified by scanner
Confidence
95% confidence
Finding

The natural-language instructions, commands, and examples are all written in Chinese, which effectively imposes a language requirement on users. There is no indication that other languages are supported or that the user can opt into Chinese, so this may violate language/locale choice expectations.

Content

No source excerpt is available for this finding.

Missing User Warnings

Low
Category
Not specified by scanner
Confidence
89% confidence
Finding

The README documents persistent writes to .milestone-state.json and storage of task details, milestone outputs, and a work session key, but provides no warning about local persistence or sensitivity of the stored data. This can expose project metadata, user task descriptions, internal progress notes, or session identifiers to other local users, backups, or source-control commits.

Content

No source excerpt is available for this finding.

Missing User Warnings

Low
Category
Not specified by scanner
Confidence
89% confidence
Finding

The skill explicitly describes persistent local state in .milestone-state.json but does not clearly warn users that task details, outputs, and session identifiers may be written to disk. In agent environments, silent persistence can expose sensitive prompts, work artifacts, or operational metadata to other local users, later runs, or backup/sync processes.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Low
Category
Not specified by scanner
Confidence
96% confidence
Finding

The file's descriptive comments and command usage/help text are written in Chinese, and all user-facing runtime messages in this script are also Chinese-only. This creates a locale constraint without any opt-in or documented justification, which matches the policy category for forced language or locale behavior.

Content

No source excerpt is available for this finding.