Back to skill

Security audit

龙虾直聘

Security checks for vulnerabilities and agentic risk

Overview

The skill is openly designed to recommend and install bundles of other skills, but it does not require enough provenance, permission, or rollback checks before changing the user's agent setup.

Review each recommended skill before installing. Confirm publisher, exact skill ID/version, requested permissions, and whether you can uninstall or roll back the bundle. Avoid using this to install unknown third-party skills into an agent environment with sensitive data or credentials.

Vulnerability Patterns
  • Insecure DependenciesIntroduces malicious components through unsafe dependency sources
  • 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)

T08 · Insecure Dependencies

Warning
Location
SKILL.md:29
Finding

Unverified Third-Party Skills May Be Installed from External Marketplaces

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, lines 29–48
Vulnerability Type: T08: Insecure Dependencies
Risk Level: Medium

Vulnerable Content

markdown
### Step 3: 在 claw.ai / ClawHub 搜索技能
- 使用公司画像里的关键词作为检索词,至少覆盖这三类角色:
`业务角色`、`交付角色`、`增长角色`。
- 每类角色至少找到 1 个候选 skill,记录:
`skill 名称`、`用途`、`适配原因`、`安装入口或标识`。
- 优先选择描述清晰、用途明确、和用户场景直接匹配的技能。

### Step 4: 产出“一键安装包”建议
- 按“基础包 + 增强包”输出,默认 3-6 个 skill:
`基础包`:保证最小可用团队能力。
`增强包`:提升效率或扩展场景。
- 明确每个 skill 对应的角色归属(如招聘官、销售、产品、工程、运营)。
- 标注安装顺序(先基础后增强)。

### Step 5: 二次确认并安装
- 在执行安装前必须显式确认,使用固定问题:
`以上是为你生成的一键安装包,是否现在开始安装?(回复:确认安装 / 调整方案)`
- 用户回复“确认安装”后再安装;若用户要求调整,先改方案再确认。
- 安装完成后输出:
`已安装`、`安装失败`、`跳过` 三类结果,并给出下一步建议。

Technical Analysis

The workflow searches external skill registries and selects candidates primarily according to descriptive clarity, stated purpose, and contextual relevance. It does not require verification of publisher identity, source repository ownership, package signatures, immutable versions, content hashes, installation scripts, requested permissions, transitive dependencies, or runtime behavior.

The explicit user-confirmation step prevents silent installation, but it does not provide an effective supply-chain security boundary. The user is not required to receive or approve security-relevant provenance, permissions, code-review results, or integrity information before confirming installation. A malicious or compromised marketplace entry could therefore be presented as a legitimate role skill based solely on convincing metadata.

The project contains no executable installation implementation, so this finding concerns the documented installation workflow rather than a confirmed malicious payload in the audited package.

Attack Path

  1. An attacker publishes a malicious, typosquatted, or compromised skill on claw.ai or ClawHub.
  2. The attacker gives the skill a clear description and role keywords that match business, delivery, or growth searches.
  3. The workflow discovers the ...[truncated 1238 chars]
Remediation
View remediation

Remediation Suggestions

  1. Restrict automatic recommendations to an explicit allowlist of trusted publishers and verified skill identifiers.
  2. Verify publisher identity and repository ownership before presenting a skill as installable.
  3. Pin every recommended skill to an immutable version, commit digest, or cryptographic package hash.
  4. Inspect the complete skill package, including instructions, scripts, installation commands, manifests, and transitive dependencies.
  5. Reject skills containing obfuscated code, remote-script execution, unexpected persistence, credential access, excessive filesystem access, or unnecessary network communication.
  6. Apply least-privilege controls and display each skill's requested permissions before confirmation.
  7. Extend the confirmation screen to show the publisher, source URL, pinned version, integrity hash, permissions, audit status, and identified security warnings.
  8. Install and test third-party skills in an isolated sandbox without production credentials or unrestricted host access.
  9. Revalidate integrity immediately before installation to prevent package replacement between review and installation.
  10. Maintain an audit log recording the selected package, publisher, immutable version, hash, permissions, confirmation event, and installation result.
Vulnerability Patterns
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
Findings (3)

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The entire file defines the taxonomy, labels, and output format exclusively in Chinese, including the required classification output fields. This imposes a specific language/locale on downstream use without any opt-in, alternative language option, or documented justification for a China-specific or Chinese-only context.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

The file hardcodes English-only output templates for user-facing interaction, including confirmation prompts, without any indication of locale detection, user preference, or opt-in. In a recruiting and team-building assistant, this can mislead or exclude users who expect Chinese or another language, increasing the risk of misunderstanding installation choices or consent-related prompts.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Low
Category
Not specified by scanner
Confidence
90% confidence
Finding

The file presents the skill name and description in Chinese while the operational default prompt is fixed in English. This imposes a language choice without any explicit opt-in or documented locale constraint, which matches the language/locale policy concern for natural-language content.

Content

No source excerpt is available for this finding.