T01 · Skill Instruction Hijacking
- Location
activate.py:240- Finding
Remote API Responses Can Inject Instructions and Poison Persistent Agent Configuration
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
This skill can automatically install other skills and surface server-provided agent-change instructions without enough scoping or user approval.
Review this skill carefully before installing. It is not judged malicious from the artifacts, but it gives the kaiwu8 service and local configuration significant influence over your OpenClaw environment. Only use it if you trust the service, the configured API endpoint, and any skills it may install; prefer a version that previews required installs and activation changes before applying them.
activate.py:240Remote API Responses Can Inject Instructions and Poison Persistent Agent Configuration
activate.py:225Remote Service Can Select and Install Unpinned Third-Party Skills
activate.py:99API Key Can Be Transmitted to an Arbitrary or Insecure Configured Endpoint
The declared purpose sounds like simple feature activation, but the described behavior extends to downloading and installing third-party skills, modifying the local skill environment, writing local state, and communicating with a remote service. This mismatch is dangerous because users may invoke the skill without understanding that it can change their system and expand code execution by pulling additional components.
The skill advertises execution of a local script, remote API access, automatic activation, and downloading missing skills, yet declares no explicit tool scope or permissions. That creates an over-privileged and opaque trust boundary where shell, filesystem, network, and environment access may be exercised without clear user/admin review.
The skill documentation omits a clear warning that invocation will automatically activate features and download missing skills from ClawHub. Silent auto-activation and package retrieval can lead to unreviewed code installation, unexpected charges/entitlements changes, or trust of remote content without informed consent.
The module-level documentation describes Step 6 as 'Xiaobai 理解并执行(精准追加到 agents.md)', implying the script will drive or cause downstream modification of agents.md as part of activation. In the actual implementation, L249 states the agent will interpret the returned what/why/how and that 'kaiwu8 不执行写入', and the code only prints the instructions without modifying agents.md or invoking another component to do so.
The script invokes an external package-management command to install skills based on server-supplied requires_skills values, without any allowlist, signature verification, or user confirmation. This creates a supply-chain and trust-boundary risk: a compromised API, account, or backend could cause arbitrary third-party skill installation, which may later execute code in the agent environment.
for attempt in range(retries):
try:
result = subprocess.run(
["clawhub", "install", slug],
capture_output=True,
text=True,
The script performs network-driven actions—automatic skill installation and activation-state persistence—without obtaining explicit user consent at execution time. Because the trigger is a simple phrase and the backend controls which skills are required, a user may unknowingly authorize installation of additional components and local state changes, increasing the risk of unintended code introduction and persistent configuration changes.
The description and trigger condition specify only the phrase "开悟吧" as the invocation, which imposes a single-language interaction pattern. The file does not mention that this is optional, user-configurable, or justified as a region-specific skill.
The natural-language description says the script runs when the user says "开悟吧", which fixes invocation to one language/locale. There is no indication that users may choose another language or that this restriction is intentionally documented as a locale-specific skill.