T01 · Skill Instruction Hijacking
- Location
assets/workspace/AGENTS.md:7- Finding
Persistent Agent Autonomy and Approval-Policy Hijacking
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
This skill is a real OpenClaw continuity installer, but it also installs broad persistent agent rules, automatic memory behavior, proactive service checks, and high-privilege example settings that deserve manual review.
Install only after reviewing the workspace files it will add. Prefer the continuity-only route, avoid --force unless you have backups and a reviewed diff, do not copy the example config with exec ask off/full security into a sensitive environment, and treat the missing runtime patch as a blocker for the full continuity route.
assets/workspace/AGENTS.md:7Persistent Agent Autonomy and Approval-Policy Hijacking
assets/workspace/AGENTS.md:56Autonomous Poisoning and Deletion of Persistent Agent Memory
assets/config/openclaw.example.json:47Example Configuration Enables Unrestricted Gateway Execution Without Approval
assets/workspace/AGENTS.md:201Continuity Package Directs Unrelated Reconnaissance of Sensitive Services
scripts/bootstrap_workspace.py:27Force Installation Can Overwrite Governing Instructions and User Context
A second description-behavior mismatch finding indicates the skill may function only as validation or anchor checking rather than as an installer or patching utility as advertised. Misrepresenting a skill that touches source trees and deployment workflows can lead users to run follow-on commands or manual steps under false assumptions, increasing the risk of unsafe edits, broken deployments, or accidental policy bypass.
A second description-behavior mismatch finding indicates the skill may function only as validation or anchor checking rather than as an installer or patching utility as advertised. Misrepresenting a skill that touches source trees and deployment workflows can lead users to run follow-on commands or manual steps under false assumptions, increasing the risk of unsafe edits, broken deployments, or accidental policy bypass.
Skill modifies its own code, configuration, or behavior at runtime. Self-modification enables an agent to escalate privileges, disable safety constraints, or install persistent backdoors.
## Why Separate?
Skills are shared. Your setup is yours. Keeping them apart means you can update skills without losing your notes, and share skills without leaking your infrastructure.
Add whatever helps you do your job. This is your cheat sheet.
Skill modifies its own code, configuration, or behavior at runtime. Self-modification enables an agent to escalate privileges, disable safety constraints, or install persistent backdoors.
default="auto",
help="Target OpenClaw workspace path. Default: auto-detect from env/current workspace, else ~/.openclaw/workspace",
)
parser.add_argument("--force", action="store_true", help="Overwrite existing files")
parser.add_argument(
"--profile",
choices=("full", "continuity"),
The skill advertises installation and patching workflows that imply shell, file read/write, and environment access, but the manifest does not declare any explicit tool scope or permissions boundary. In an agent ecosystem, missing scope declarations can cause operators and automated policy layers to underestimate what the skill may do, increasing the chance of unintended source modification or command execution.
The skill references a full continuity install path that applies a runtime patch, rebuilds, and redeploys against a matching source tree, but it does not provide a prominent upfront warning about these source modification and deployment-impacting actions. Users may invoke destructive or environment-altering steps without understanding that application code will be changed and rebuilt, which is especially risky in production-like environments.
The documentation states that the runtime patch enables silent continuity preparation and suppresses visible continuity/context hint messages, which changes user-visible chat behavior, but it does not frame this as a clear behavioral-risk warning. Hidden or less visible system behavior can affect user expectations, transparency, debugging, and auditability, particularly where conversation continuity is being preserved without explicit prompts.
The manifest text uses very broad activation language such as '全自动新对话续接' and '一键安装' without clearly constraining when the skill should be invoked or what user consent is required. Combined with allow_implicit_invocation: true, this increases the chance the agent will trigger the skill in unrelated contexts and perform installation-oriented actions or continuity patch guidance the user did not explicitly request.
The visible metadata and default behavior are written only in Chinese and do not indicate any user language negotiation or documented locale restriction. This can cause users to misunderstand what the skill will do, especially because it discusses installation paths and runtime patch behavior, reducing informed consent and increasing the risk of accidental invocation or misuse.
This configuration enables automatic memory flushing that silently writes durable state to workspace files and instructs the agent to reply with NO_REPLY unless absolutely necessary. That creates a covert persistence channel for conversation-derived data, increasing the risk of storing sensitive information without user awareness or meaningful consent.
The elevated-tool activation scope is defined with generic placeholders for both channel and sender, which provides no concrete security boundary in the example configuration. In a skill that explicitly enables elevated tools, an operator may replace these loosely or incorrectly, causing privileged actions to be granted to unintended message sources and enabling command execution or other sensitive operations.
The exec tool is configured with full security, hosted on the gateway, and prompting turned off, allowing shell or execution actions to proceed without interactive approval. In the context of a continuity/automation skill that aims for silent background continuation, this materially increases the chance of unreviewed command execution, misuse of credentials, and destructive system actions.
The instruction not to ask the user for harmless reads or dry-runs can be acceptable in tightly sandboxed local environments, but as a reusable skill it weakens consent boundaries and normalizes silent tool execution. Combined with the document's broader autonomy themes, it may cause an agent to act with insufficient confirmation about environment, scope, or data sensitivity.
- what remains
- the only blocker, if any
7. Never create or modify AGENTS.md again just because the user repeats this preference. Treat this rule as persistent.
8. If this local Control UI session already has exec configured as ask=off and security=full, do not ask the user to paste /approve for harmless local reads or dry-runs. Execute them directly and only mention /approve when the tool explicitly returns approval-pending.
## First Run
The activation condition says to treat any task with more than one phase, more than one workstream, or likely to take more than about 10 minutes as a long task. Those criteria are broad and subjective, and the file does not provide negative examples or tighter scope boundaries, which could cause unintended invocation of the long-task workflow.
Skill enables autonomous high-impact decisions without human-in-the-loop verification. Critical operations (destructive commands, financial transactions, data deletion) should require explicit user confirmation.
## Red Lines
- Don't exfiltrate private data. Ever.
- Don't run destructive commands without asking.
- `trash` > `rm` (recoverable beats gone forever)
- When in doubt, ask.
The heartbeat section authorizes proactive checks of external services and even 'commit and push your own changes' without a strong per-action consent boundary. That expands the agent from reactive assistant to autonomous actor with outbound network behavior, creating risk of unintended disclosure, unauthorized communications, or repository changes performed outside explicit user intent.
The instruction explicitly permits proactive work 'without asking,' including checking projects and committing/pushing changes. In skill context, this is more dangerous because the same document also encourages heartbeat-driven background activity and external checks, increasing the chance of silent modifications or outbound actions that the user did not explicitly request in the current session.
- Nothing new since last check
- You just checked <30 minutes ago
**Proactive work you can do without asking:**
- Read and organize memory files
- Check on projects (git status, etc.)
The activation criteria are broad and subjective (for example, 'likely to touch multiple files/systems' or 'anything the user may want resumed later'), which can cause the protocol to trigger on ordinary work without clear user intent. In this skill's context, that increases the chance of unnecessary persistence and continuity-file creation, expanding data retention and making subsequent automated behavior less predictable.
The protocol directs the agent to automatically append and create/update memory, plan, and status files immediately on task intake, but provides no requirement to notify the user or obtain consent. In a continuity skill, this is especially risky because it silently persists task details and decisions to disk, which may capture sensitive prompts, operational context, or user data beyond the user's expectations.
The file's core instruction is written as a direct persona/specification in Chinese, telling the assistant what it is, but it does not indicate that Chinese is optional or user-selected. This creates a language/locale policy concern because the skill appears to impose a specific language context without opt-in or justification.
The file forces a specific language for operational instructions and does not provide an opt-in, alternative language, or justification for the locale restriction. Under the stated policy, fixed language behavior without user choice or documented rationale is a natural-language policy violation.
This markdown file uses Chinese throughout and does not provide any indication that the user can choose another language or that the document is intentionally limited to a Chinese-speaking audience. Under the stated policy, forcing a specific language without opt-in is a natural-language policy violation.
This markdown file uses only Chinese for critical installation and deployment guidance, which can amount to a language policy violation when no user opt-in or alternative locale is offered. The document does not indicate that the skill is China-specific or otherwise justified to require Chinese-only instructions.
The document begins entirely in Chinese and presents the skill usage instructions in that language, with no indication that other languages are supported or that the language choice is optional. Under the policy, forcing a specific language without user opt-in is a natural-language policy violation unless clearly justified as region-specific.
This markdown file contains operational instructions primarily in Chinese, but also includes English-only prompt examples and interface terminology such as 'Please reply with READY only.' and 'Control UI'. That effectively imposes a mixed language/locale expectation without any opt-in or explanation, which can conflict with organizational language-choice policies.