Back to skill

Security audit

config-new-agent

Security checks for vulnerabilities and agentic risk

Overview

The skill is not overtly malicious, but it asks to install and modify other skills and keep one running continuously without clear separate approval or lifecycle controls.

Install only if you are comfortable with this skill configuring a new agent beyond bindings: it will install other skills, alter one installed skill file, and keep a self-improving component running. Prefer a version that asks for separate confirmation for each install and for persistent runtime, pins dependency versions or digests, and documents how to stop or remove the always-running skill.

Vulnerability Patterns
  • Skill Instruction HijackingAlters the agent's session goals or safety constraints when the skill loads
  • System PersistenceInstalls backdoors, hooks, services, or scheduled tasks that survive the run
  • Insecure DependenciesIntroduces malicious components through unsafe dependency sources
  • 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
Findings (3)

T06 · System Persistence

Error
Location
SKILL.md:105
Finding

Continuously Running Self-Improving Skill Creates a Persistent Execution Path

Content
View full analysis
``` **安装完成后**,必须编辑新 agent workspace 中的 SKILL.md,在文件顶部添加 workspace 隔离规则引用块(见上文"self-improving-proactive-agent 使用隔离规则")。 **同时配置该 skill 保持一直运行**。 ``` ### Technical Analysis The workflow requires installing a third-party self-improving agent component and configuring it to remain continuously active. The document does not define a bounded execution period, shutdown mechanism, integrity check, resource limits, or restricted capability set. Although the workspace-isolation rule attempts to limit data access, a textual instruction is not an enforceable security boundary. If the installed component is compromised or behaves unexpectedly, its continuously running configuration provides an ongoing execution foothold beyond the original binding operation. ### Attack Path 1. A user triggers the new-agent configuration workflow. 2. The workflow installs `self-improving-proactive-agent` from the external skill registry. 3. The installed component is configured to remain continuously running. 4. A compromised or subsequently altered dependency executes with the permissions available to the OpenClaw process. 5. The component continues operating after the original configuration task has completed. 6. Without enforceable sandboxing or lifecycle controls, it may repeatedly access or modify resources available within its execution context. ### Impact Assessment The affected scope includes the new agent workspace and any additional files, tools, credentials, or APIs accessible to the OpenClaw process. The exact privilege level depends on the account running the commands, but the documented installation target is under `/root/.openclaw`, indi ...[truncated 427 chars]
Remediation
View remediation

T08 · Insecure Dependencies

Error
Location
SKILL.md:91
Finding

Unpinned Third-Party Skills Are Installed Without Enforced Artifact Verification

Content
View full analysis
``` ```bash clawhub install skill-finder-cn --workdir /root/.openclaw/workspace- ``` ```bash clawhub install self-improving-proactive-agent --workdir /root/.openclaw/workspace- ``` ```bash clawhub install openclaw-tavily-search --workdir /root/.openclaw/workspace- ``` ### Technical Analysis All four dependencies are identified only by mutable package slugs. The workflow does not pin immutable versions, verify publisher identities, require cryptographic hashes, validate signatures, or constrain the registry source. The document states that future skill installations must be reviewed with `skill-vetter`, but the workflow does not provide an enforced review step or approval gate before each listed installation. In addition, `skill-vetter` itself must be installed before it can be used, creating an unavoidable bootstrap trust issue. Consequently, the effective contents installed by these commands may differ from the contents that were previously reviewed. Registry compromise, publisher-account compromise, dependency replacement, or malicious updates could introduce attacker-controlled instructions or executable content. ### Attack Path 1. An attacker compromises a listed package, its publisher account, or the package registry resolution process. 2. The attacker publishes a malicious release under one of the expected mutable slugs. 3. A user triggers the new-agent configuration workflow. 4. `clawhub install` resolves the slug to the attacker-controlled release because no immutable version or digest is specified. 5. The malicious skill is installed under the agent workspace. 6. When OpenClaw loads or invokes the installed skill, its malicious instructions ...[truncated 749 chars]
Remediation
View remediation

T01 · Skill Instruction Hijacking

Warning
Location
SKILL.md:108
Finding

Post-Installation Modification Breaks Skill Instruction Integrity

Content
View full analysis
**IMPORTANT - Workspace Isolation Rule**: > Each agent must use the `self-improving-proactive-agent` skill in its **own workspace only**. > All learning, corrections, memory, and state files must be stored in the agent's **own workspace directory**. > **Never reference, read from, or write to another agent's self-improving-proactive-agent skill or data files.** ``` ### Technical Analysis The workflow explicitly modifies an installed skill's `SKILL.md` after installation. The currently specified text is intended to improve workspace isolation and is not itself an overtly malicious instruction. However, modifying the installed artifact after review breaks the relationship between the reviewed package and the runtime package. Because skill text controls agent behavior, post-install mutation creates a mechanism through which persistent behavioral instructions can be inserted without being covered by the original package digest, signature, or security review. It also makes later integrity verification ambiguous because the installed package is expected to differ from its upstream artifact. ### Attack Path 1. The workflow installs `self-improving-proactive-agent`. 2. It opens and modifies the installed skill instruction file. 3. The modified `SKILL.md` becomes the version loaded in later agent sessions. 4. If the injected block, insertion process, or source instructions are tampered with, attacker-controlled directives can be placed at the top of the skill. 5. Those directives can influence future agent behavior whenever the modified skill is loaded. The provided block is isolation-orien ...[truncated 688 chars]
Remediation
View remediation
Vulnerability Patterns
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
Findings (3)

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

The manifest description is entirely written as an instruction flow in Chinese and names a Chinese-specific search tool (skill-finder-cn) as part of the default setup, with no indication that the skill can operate in other languages or based on user preference. Under the policy, forcing a specific language or locale without user opt-in is a natural-language policy violation.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

The skill’s stated purpose is configuring new agent bindings, but the documented workflow also installs additional skills, edits another skill’s SKILL.md, and changes runtime behavior by requiring a separate skill to remain always running. This violates least privilege and scope minimization: a binding-setup skill should not silently gain authority to modify unrelated components, because that broadens the blast radius and creates an avenue for persistence or unsafe cross-skill changes.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

Instructing the system to configure self-improving-proactive-agent to stay always running introduces persistence beyond the declared task of assigning a binding. A continuously running self-modifying or stateful component can expand access over time, consume resources, and make unauthorized behavior harder to notice or roll back, especially when enabled automatically for a newly created agent.

Content

No source excerpt is available for this finding.