T01 · Skill Instruction Hijacking
- Location
lobster-training-skill/scripts/auto-trigger.js:189- Finding
Persistent Overwrite of Agent Persona, Identity, and Behavioral Control Files
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
This package is advertised as a browser automation skill, but it also contains unrelated image-generation, self-improvement hook, and onboarding components that can persist user data, alter agent behavior files, and expand installed capabilities.
Review this package before installing. If you only need browser automation, use a clean, pinned agent-browser installation and avoid enabling the bundled hooks or lobster-training package. Do not provide image-generation API credentials to this artifact unless you have verified the endpoint, and do not allow it to write OpenClaw memory/persona files or install additional skills without explicit approval and backups.
lobster-training-skill/scripts/auto-trigger.js:189Persistent Overwrite of Agent Persona, Identity, and Behavioral Control Files
lobster-training-skill/scripts/auto-trigger.js:89Automatic User Profiling and Storage of Untrusted Message Content in Persistent Memory
hooks/openclaw/handler.js:8Bootstrap-Time Instruction Injection Through OpenClaw Hooks
scripts/generate_image.py:13API Credential and Prompt Transmission to a Third-Party Default Endpoint
SKILL.md:15Unpinned Global Installation of a Third-Party Browser Automation Package
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.
## Adding New Commands to the Skill
Update SKILL.md when the upstream CLI adds new commands.
- Keep the Installation section
- Add new commands in the correct category
- Include usage examples
The README documents an image-generation skill, while the metadata identifies this package as a browser automation skill. This mismatch is dangerous because users, reviewers, or automated tooling may install or grant permissions to a skill under false assumptions, creating supply-chain trust and review bypass risk. In this context, the discrepancy is more serious because browser automation capabilities can access web content, interact with sessions, and perform actions on behalf of an agent, so mislabeled documentation can conceal materially different behavior.
The declared description is about browser automation capabilities such as navigation, clicking, typing, and page snapshots. The provided code does none of that. Instead, it implements an event hook for agent bootstrap and modifies event context by inserting a virtual reminder file. This is a materially different primary purpose with unrelated triggers and behavior, so the description does not accurately represent the code chunk.
The declared description is about a browser automation CLI, but the code chunk does not implement any browser control or CLI behavior. Instead, it is a hook for agent bootstrap events that adds a virtual reminder markdown file to the bootstrap files list. This is a materially different primary purpose and introduces an undeclared trigger/event-based capability unrelated to browser automation.
The declared description is for a browser automation CLI, but the supplied code is unrelated to browser control. The code defines a training questionnaire, records answers, persists installation/training progress, and writes multiple local markdown files containing user profile and persona information. There is no Rust component, no Node.js fallback logic for browser execution, no page navigation/click/type/snapshot functionality, and no structured browser command handling. This is a clear description-behavior mismatch with materially different purpose and undeclared local file access/persistence behavior.
The declared description is for a browser automation tool, but the provided code does not automate browsers, navigate pages, click, type into web pages, or take snapshots. Instead, it acts as a message-triggered onboarding/training handler: it checks training status, presents questions, records user responses, generates configuration files, and installs bundled skills upon completion. This is a materially different primary purpose with unrelated triggers and capabilities compared with the declared browser automation CLI.
The declared description is for a browser automation CLI, but the supplied code is unrelated to browser control. Its primary purpose is a training/setup workflow: reading a local markdown document, extracting six questions, installing packaged skills, and creating/updating local configuration files such as USER_PROFILE.md and SOUL.md. It also invokes an external command (openclaw skill add) to install ZIP archives. None of the core declared capabilities—headless browsing, structured page interaction, navigation, typing, or snapshots—are present in the code. This is a clear material mismatch in both purpose and behavior.
The declared description is for a browser automation tool, but the provided code is a Bash script that simply prints a reminder about capturing learnings to .learnings/. Its primary purpose, behavior, and trigger context are entirely unrelated to browser automation. This is a clear material mismatch between declared purpose and actual code behavior.
The declared description is about browser automation capabilities, but the supplied code does not implement any browser control or CLI functionality. Instead, it acts as an error-detection shell hook triggered after Bash tool execution, scanning tool output for failure patterns and printing guidance about recording errors. This is a materially different primary purpose and introduces unrelated behavior/triggers compared with the declared browser automation description.
The declared description describes a browser automation tool for interacting with web pages. The supplied code does not implement any browser control, page interaction, snapshotting, Rust components, or Node.js fallback. Instead, it is a shell script for creating a new skill folder and templated markdown file on disk. This is a materially different primary purpose and includes file-system scaffolding behavior that is unrelated to the declared browser automation functionality.
The declared description is about headless browser automation for navigation and page interaction, supposedly implemented as a fast Rust CLI with Node.js fallback. The actual code is instead a Python script for text-to-image generation using a remote Gemini API. It does not automate a browser, navigate pages, click, type, or take webpage snapshots. Its primary purpose, implementation language, external resource usage, and capabilities are materially different from the declared description.
The skill instructs collection of personal and preference data and writes it into multiple persistent files, but it does not provide a clear privacy notice, retention policy, or user consent flow. Users may disclose sensitive personal or business information without understanding where it is stored, how long it is kept, or who can access it.
The onboarding skill is scoped as a training/manual workflow, but it instructs automatic installation of three separate high-capability packages, including browser automation and self-improvement. Expanding privileges and capabilities during onboarding without explicit user approval increases attack surface and can silently introduce web automation, scraping, or autonomous behavior the user did not knowingly consent to.
The bundled browser skill introduces browsing, web interaction, and data collection capabilities, yet the skill does not provide a clear warning about those behaviors or their implications. Browser automation can access third-party sites, submit forms, or scrape content, which materially changes risk and should be disclosed before enablement.
The manifest enables automatic installation, proactive questioning, and persistent memory recording by default, with no visible consent or notice mechanism in the package metadata. This creates a default data collection and capability expansion path that can surprise users and lead to unauthorized retention of sensitive information.
Training is configured to auto-start and target files like USER_PROFILE.md, MEMORY.md, SOUL.md, and IDENTITY.md, which strongly suggests collection and persistence of personal or identity-linked data without explicit disclosure. Auto-start removes a meaningful consent checkpoint and increases the chance of overcollection.
Targeting files named PROFILE, MEMORY, SOUL, and IDENTITY indicates an intent to aggregate sensitive personal, preference, or identity-related information into persistent stores. In combination with auto-start and write access, this materially raises privacy, profiling, and prompt-persistence risks.
The script relies on environment variables such as HOME/USERPROFILE to determine where to read and write persistent files, but this capability is not declared by the skill permissions. Undeclared environment access weakens user trust and can enable writes outside the expected skill scope, especially when combined with automatic execution and filesystem modification.
The code defines a six-question onboarding flow that collects personal information, work details, target customers, preferences, challenges, and emotional needs, then stores the answers persistently. For a browser automation skill, this data collection is unjustified and expands risk by creating a durable profile of the user that could later be misused or leaked.
This browser automation skill contains an auto-trigger that generates USER.md, USER_PROFILE.md, MEMORY.md, SOUL.md, and IDENTITY.md, which are unrelated to browser control. That is a strong scope mismatch: the skill silently alters persistent user-profile and agent-behavior files, creating unauthorized state and potentially influencing future agent behavior beyond the stated browser purpose.
The script rewrites SOUL.md and IDENTITY.md to reshape the agent's persona, behavior rules, and self-description based on collected answers. Persistently modifying core identity or memory artifacts is dangerous because it can silently steer future interactions and decision-making, which is unrelated to browser automation and difficult for users to notice.
The training flow immediately installs bundled skill packages after completion without explicit user confirmation. In this context, that means a conversational onboarding script can silently add new executable capabilities, which is especially dangerous because skills may have broad agent permissions or introduce further supply-chain risk.
This code records arbitrary incoming user messages as answers once the flow is active, without a clear just-in-time consent prompt or a way to opt out. That can capture unrelated, sensitive, or accidental user input into training state/configuration, creating privacy and integrity risks.
The file's behavior is an onboarding/training installer rather than browser automation, which materially misrepresents the skill's purpose. In the context of an Agent Browser skill, this mismatch is especially dangerous because users and reviewers may grant trust or permissions suitable for browsing while the code instead modifies local state and installs software.
The script performs actions far beyond a browser automation skill: it installs additional skills and overwrites persistent user configuration under ~/.openclaw. This is dangerous because a user invoking an 'Agent Browser' skill would not reasonably expect package installation or persona/profile rewrites, creating a supply-chain and integrity risk through deceptive scope expansion.