Back to skill

Security audit

哈啰顺风车出行技能

Security checks for vulnerabilities and agentic risk

Overview

This ride-booking skill is mostly purpose-aligned, but it gives agents under-scoped authority to schedule future automatic ride orders and run cron commands for real-world transactions.

Review carefully before installing. Use a dedicated, revocable API key, avoid pasting secrets into chat, and do not enable the cron-based automatic booking path unless you are comfortable with future background tasks creating real ride orders. Prefer interactive booking where the agent shows the exact route, price, capacity, wait time, and irreversible trip action before you confirm.

Vulnerability Patterns
  • Insecure Skill Coding PracticesFinds exploitable flaws such as hardcoded secrets or command injection
  • 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 (4)

T09 · Insecure Skill Coding Practices

Error
Location
SKILL.md:305
Finding

Shell Command Injection Through Unsanitized Cron Parameters

Content
View full analysis

Vulnerability Details

File Location: SKILL.md:305-325
Vulnerability Type: Shell command injection
Risk Level: High

Vulnerable Code

bash
openclaw cron add \
  --name "hitch-reserve:$(date +%s)" \
  --at "TIME_VALUE" \
  --session isolated \
  --message "[Automatically book a ride from ORIGIN to DESTINATION and create a follow-up task]" \
  --announce \
  --channel CHANNEL_NAME \
  --to "CHAT_ID"

A second template uses the same unsafe construction:

bash
openclaw cron add \
  --name "hitch-status:ORDER_GUID" \
  --at "5m" \
  --session isolated \
  --message "[Query the latest state of order ORDER_GUID]" \
  --announce \
  --channel CHANNEL_NAME \
  --to "CHAT_ID"

The message text above is an English rendering of the source text; the command structure and dynamic placeholders are unchanged.

Technical Analysis

The templates construct shell commands using values derived from ride addresses, order identifiers, and conversation metadata. The instructions do not require validation or shell-safe encoding before substituting ORIGIN, DESTINATION, ORDER_GUID, CHANNEL_NAME, or CHAT_ID.

CHANNEL_NAME is especially dangerous because it is not quoted. Although CHAT_ID and the message are enclosed in double quotes, double quotes do not prevent shell command substitution such as $(command) or backtick substitution. A malicious value could therefore alter command arguments or cause the shell to execute a nested command.

The vulnerability applies when an Agent implements the documented template through textual substitution and passes the result to a shell.

Attack Path

  1. An attacker supplies a crafted address or influences a session metadata value.
  2. The value contains shell syntax, such as command substitution or argument-breaking metacharacters.
  3. The Agent substitutes the value into the documented cron command.
  4. The resulting string is ...[truncated 675 chars]
Remediation
View remediation

Remediation Suggestions

  • Do not construct commands through string interpolation or invoke them through a shell.
  • Launch openclaw with a structured process API that accepts an executable and an argument array.
  • Pass the task message, channel, destination, and schedule as separate arguments.
  • Apply strict allowlists to CHANNEL_NAME and validate identifiers against documented formats.
  • Reject control characters, shell metacharacters, command-substitution syntax, and unexpected line breaks.
  • If shell execution is unavoidable, apply a proven shell-escaping routine independently to every dynamic value.
  • Prefer a native cron-management tool or MCP API over documented shell templates.
  • Run scheduling operations under a minimally privileged account with restricted filesystem and network access.

T09 · Insecure Skill Coding Practices

Error
Location
SKILL.md:305
Finding

Scheduled Ride Booking Bypasses Required Price and Capacity Approval

Content
View full analysis

Vulnerability Details

File Location: SKILL.md:305-313
Vulnerability Type: Transaction confirmation bypass
Risk Level: High

Vulnerable Code

bash
openclaw cron add \
  --name "hitch-reserve:$(date +%s)" \
  --at "TIME_VALUE" \
  --session isolated \
  --message "[Automatically book a ride from ORIGIN to DESTINATION. Resolve the addresses, obtain a quotation, and create the order. After successful creation, report the order and create a five-minute status-check task.]" \
  --announce \
  --channel CHANNEL_NAME \
  --to "CHAT_ID"

The source also describes the resulting behavior as:

text
The cron task triggers and automatically completes search, quotation, and order creation.

These snippets are faithful English renderings of the source instructions.

Technical Analysis

The normal booking workflow requires the Agent to:

  1. Resolve and, when necessary, confirm both addresses.
  2. Obtain the current quotation.
  3. Display all available capacity and price options.
  4. Wait for the user to select one or more options.
  5. Confirm the permitted waiting duration.
  6. Create the order.

The isolated scheduled-task instruction instead directs a new Agent session to resolve addresses, obtain a quotation, and create the order automatically. It does not specify a previously approved capacity selection, a maximum acceptable price, an exact resolved location, or a mechanism to pause for confirmation.

Because the scheduled session has no conversation history, it cannot reliably establish that the user approved the actual price and options returned at execution time. This creates an authorization and business-logic gap between the interactive workflow and the scheduled workflow.

Attack Path

  1. A future ride task is created based on a general request from the user.
  2. At the scheduled time, an isolated Agent session resolves the locations and requests a new quotation.

...[truncated 774 chars]

Remediation
View remediation

Remediation Suggestions

  • Do not let a scheduled session independently choose a capacity or accept an unbounded price.
  • Before scheduling, resolve and explicitly confirm the exact origin, destination, departure time, capacity, waiting duration, and maximum acceptable price.
  • Store the approved structured parameters rather than a natural-language booking instruction.
  • At execution time, compare the current quotation with the approved limits.
  • If the price, capacity, route, or resolved location differs, stop and request fresh user approval.
  • Prefer planDepartureTime after explicit confirmation of all booking terms instead of a separate shell-based Agent session.
  • If interactive approval is unavailable at execution time, make the scheduled task quotation-only and prohibit order creation.
  • Record an auditable confirmation token tied to the exact route and transaction constraints.

T09 · Insecure Skill Coding Practices

Warning
Location
SKILL.md:269
Finding

Irreversible Trip-State Operations Lack Explicit Confirmation Requirements

Content
View full analysis

Vulnerability Details

File Location: SKILL.md:269-271
Vulnerability Type: Unsafe irreversible operation handling
Risk Level: Medium

Vulnerable Code

text
Confirm boarding: call hitch_pax_confirm_get_on_car.

Confirm arrival: call hitch_pax_confirm_reach_destination.

The corresponding API documentation states:

text
Passenger confirmation of boarding: irreversible.
Passenger confirmation of arrival at the destination: irreversible.

These snippets are faithful English renderings of SKILL.md:269-271 and references/api.md:263-286.

Technical Analysis

The Skill explicitly requires secondary confirmation for cancellation and requires user selection before inviting a driver. It does not impose an equivalent confirmation step immediately before boarding or arrival calls, even though the API documentation identifies both operations as irreversible.

Natural-language phrases such as “I am here,” “we arrived,” or “the driver is here” can be contextually ambiguous. Without an explicit confirmation requirement tied to a specific order, an Agent could interpret such a phrase as authorization to change the trip state.

Idempotency keys prevent duplicate execution of the same operation, but they do not establish user intent and cannot reverse an incorrectly authorized state transition.

Attack Path

  1. A ride order is active.
  2. The user sends an ambiguous statement concerning arrival or the driver's location.
  3. The Agent classifies it as a request to confirm boarding or destination arrival.
  4. The Agent immediately invokes the corresponding API because the instructions do not require a confirmation prompt.
  5. The platform records an irreversible trip-state transition.

Impact Assessment

Premature boarding confirmation may incorrectly begin the trip. Premature arrival confirmation may complete fulfillment or affect settlement before the passenger reaches the destinatio ...[truncated 324 chars]

Remediation
View remediation

Remediation Suggestions

  • Require an explicit confirmation immediately before either irreversible API call.
  • Display the order identifier, route, current state, and proposed transition in the confirmation prompt.
  • Accept only an unambiguous affirmative response referring to that action and order.
  • Do not treat generic statements such as “arrived” as sufficient confirmation when multiple interpretations or orders are possible.
  • Query the current order state before executing the transition and reject invalid state changes.
  • Expire pending confirmations after a short interval or after any material order-state change.
  • Preserve idempotency controls, but do not treat idempotency as a substitute for authorization.

T09 · Insecure Skill Coding Practices

Warning
Location
README.md:12
Finding

Documentation Encourages Disclosure of API Credentials in Agent Chat

Content
View full analysis

Vulnerability Details

File Location: README.md:12-15
Vulnerability Type: Insecure credential handling
Risk Level: Medium

Vulnerable Code

text
2. Configure the integration by saying this in the Agent conversation:

User: My API Key is xxxxxx

This is a faithful English rendering of the source example.

Technical Analysis

The quick-start instructions encourage users to submit an API key through an ordinary Agent conversation. Chat messages may be stored in conversation history, observability systems, gateway logs, model-provider records, support exports, or screenshots. They may also remain available in the active model context.

API keys are bearer credentials. Anyone obtaining the value may be able to invoke the configured ride-service API within the permissions and quotas associated with the key. The repository also provides a local MCP configuration method, making disclosure through chat unnecessary.

Attack Path

  1. A user follows the quick-start documentation and pastes the API key into the conversation.
  2. The message is retained by one or more chat, Agent, gateway, or logging components.
  3. A party with access to those records obtains the plaintext key.
  4. The party configures its own client with the exposed key.
  5. The key is used to invoke API operations until it is revoked, expires, or is otherwise restricted.

Impact Assessment

Exposure may permit unauthorized use of Hellobike MCP operations associated with the credential, consumption of API quotas, access to order information available to the key, and initiation or modification of ride workflows where the backend permits it.

The exact scope depends on server-side authorization, account binding, credential permissions, and transaction confirmation controls.

Remediation
View remediation

Remediation Suggestions

  • Remove the instruction to paste an API key into an Agent conversation.
  • Direct users to a protected local configuration interface, environment variable, operating-system keychain, or dedicated secret manager.
  • Ensure configuration commands do not expose secrets through shell history or process listings.
  • Mask credentials in diagnostic output and redact authorization headers from logs.
  • Prevent secret values from entering model context.
  • Document key rotation and immediate revocation procedures.
  • Recommend narrowly scoped, short-lived credentials where supported.
  • Add automated secret detection and redaction at conversation and gateway boundaries as defense in depth.
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
Findings (12)

Vague Triggers

High
Category
Not specified by scanner
Confidence
96% confidence
Finding

The top-level description includes broad natural-language triggers such as everyday travel phrases, which can cause accidental activation from normal conversation. In a skill that can progress toward querying prices, generating links, or creating orders, overbroad activation increases the chance of unintended real-world actions and collection of location-sensitive data.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

High
Category
Not specified by scanner
Confidence
95% confidence
Finding

The documentation contains shell-command-based cron management instructions, including dynamic command construction and execution guidance for scheduling future actions. Embedding system-level task management in a ride-booking skill creates a capability expansion path where an agent may perform host-side operations unrelated to core ride assistance, increasing the blast radius if the skill is misused or the agent over-trusts the instructions.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The README instructs users to provide an API key in plain agent conversation and shows storing it directly in configuration without any warning about secret handling. In an agent ecosystem, chat transcripts, logs, screenshots, shell history, or shared config files may expose the credential, enabling unauthorized use of the ride-ordering account and associated APIs.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
87% confidence
Finding

The skill handles sensitive trip data including real-time driver location, home addresses, and ride booking/cancellation actions, but the documentation lacks any privacy or safety warning. This can normalize unsafe use around highly sensitive mobility data, increasing the risk of unintended disclosure, stalking, account misuse, or accidental real-world actions.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
78% confidence
Finding

The skill description, triggers, examples, and user-facing phrasing are entirely in Chinese, and no opt-in or alternative language behavior is offered. This can amount to a language policy issue if the skill is expected to support users without implicitly forcing a specific language.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

The scenario table uses generic triggers like '回家', '上班', and similar short phrases that may appear in ordinary chat without an actual request to invoke the skill. Because the skill handles real transportation workflows, ambiguous activation raises the risk of premature location inference, order-flow initiation, or confusion about whether the user intended a transactional action.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

The skill instructs the agent to create delayed cron tasks that will later perform real-world actions such as querying order state or placing a ride order in an isolated session. This extends the skill from conversational assistance into autonomous persistence and scheduled execution, which increases risk because actions can occur outside the user's immediate oversight and after context has changed.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
89% confidence
Finding

The skill describes automatic order creation through scheduled tasks without a strong user-facing warning that a future autonomous process may place a real ride request and potentially incur charges. This is dangerous because the action may execute when the user is unavailable, circumstances have changed, or the user did not understand that a background task would perform a live transactional operation.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The API documents irreversible passenger actions ('confirm get on car' and 'confirm reach destination') as non-revocable, but unlike order creation/cancellation/invite operations it does not explicitly require user confirmation before invocation. In a ride-hailing assistant, an LLM or integration bug could trigger these state transitions without clear user consent, causing premature trip progression, billing disputes, or loss of ability to correct the order state.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
80% confidence
Finding

The operational instructions, example prompts, and required reply texts are entirely specified in Chinese, including mandated response phrasing such as the exact refusal and user-facing templates. There is no indication that the user can choose another language or that the locale restriction is explicitly justified as a region-specific requirement.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
90% confidence
Finding

The manual exposes a direct capability to query and display a driver's real-time location but does not document any user-visible warning, consent boundary, minimization rule, or authorization check before doing so. In a ride-hailing context this is privacy-sensitive location data about another party, and insufficient guardrails can enable overexposure, stalking, or disclosure to someone who should not have access to the trip.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Low
Category
Not specified by scanner
Confidence
76% confidence
Finding

The README is entirely in Chinese and all example interactions are presented only in Chinese, with no indication that users may choose another language. This can constitute a language/locale policy issue unless the locale restriction is explicitly documented and justified as a region-specific requirement.

Content

No source excerpt is available for this finding.