Back to skill

Security audit

Hyderabad Shopper

Security checks for vulnerabilities and agentic risk

Overview

This skill is a shopping helper, but it goes beyond price comparison by preparing checkout and sending checkout screenshots over Telegram without clear upfront disclosure.

Review this skill carefully before installing. It may be acceptable only if you explicitly want a shopping assistant that can prepare carts and message you externally, but the current artifact should disclose that behavior, avoid sending checkout screenshots unless specifically approved, and remove the hardcoded full address from the package.

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 (2)

T09 · Insecure Skill Coding Practices

Warning
Location
SKILL.md:5
Finding

Hardcoded Precise Residential Location in Distributable Skill Metadata

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, lines 5–7
Vulnerability Type: Plaintext sensitive personal data exposure
Risk Level: Medium

Vulnerable Code:

yaml
metadata:
  address: "102 Saboori Enclave Whitefields"
  pincode: "500081"
  location: "Hyderabad"

Technical Analysis

The skill stores a precise residential address directly in plaintext metadata. This information is available to anyone who can access, distribute, inspect, archive, or index the skill package. The workflow only needs a delivery pincode to search for product availability, so retaining the complete address violates data-minimization principles.

No access control, runtime prompt, environment-based configuration, or redaction mechanism protects the address. Because it is static package content, deleting the address from an individual runtime session would not remove it from existing copies, repositories, archives, or caches.

Attack Path

  1. An attacker obtains access to the skill package or a repository containing it.
  2. The attacker opens SKILL.md without needing to execute the skill.
  3. The attacker reads the metadata section and extracts the precise address, pincode, and city.
  4. The attacker correlates that information with other available records or uses it for profiling, phishing, social engineering, or physical-location targeting.

Impact Assessment

Exploitation requires only read access to the distributed project and grants no direct system privileges. However, it exposes persistent personal location information beyond what is necessary for the skill's functionality. The disclosure scope includes every person or service that receives, stores, logs, indexes, or republishes the skill package.

Remediation
View remediation

Remediation Suggestions

  • Remove the full street address from the skill metadata.
  • Request the minimum required location information at runtime, preferably only the pincode needed for product availability checks.
  • Store user-specific location data outside distributable skill files and prevent it from being committed to source control.
  • If persistent location configuration is essential, use access-controlled local configuration and document its retention policy.
  • Review repository history, published packages, caches, and backups for previous copies of the exposed address and remove them where feasible.
  • Add automated secret and sensitive-data scanning for addresses, phone numbers, account identifiers, and other personal information before publication.

T09 · Insecure Skill Coding Practices

Warning
Location
SKILL.md:35
Finding

Mandatory Transmission of Potentially Sensitive Checkout Screenshots via Telegram

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, lines 35–42
Vulnerability Type: Unsafe external disclosure of checkout information
Risk Level: Medium

Vulnerable Code:

markdown
## Phase 5: Human-in-the-Loop Approval (CRITICAL)
1. **PAUSE EXECUTION.** Do not click "Pay Now," "Place Order," or any payment confirmation button.
2. Send a screenshot of the checkout page to the user via Telegram.
3. Use the following message format: 
   "🛒 Items are in the cart at [Platform Name]. 
    Total: ₹[Amount] (Delivery to 500081). 
    **Please check the browser window on your Mac to complete the payment.**"
4. Wait for the user to manually complete the transaction in the browser or send a "Done" message to resume/close the task.

Technical Analysis

The skill mandates capturing a checkout page and transmitting the resulting screenshot through Telegram. Checkout pages can display recipient names, delivery addresses, telephone details, account identifiers, cart contents, order totals, loyalty information, and partial payment details.

The instructions do not require the agent to inspect or redact the screenshot, obtain explicit consent for external transmission, verify the Telegram destination, or confirm that the screenshot excludes payment and identity data. This creates a privacy and information-disclosure risk even though the skill correctly prohibits clicking the final payment confirmation button.

Telegram transmission also expands the data's exposure beyond the local browser to messaging infrastructure, chat history, synchronized devices, notification previews, backups, and any incorrectly configured or compromised recipient account.

Attack Path

  1. A user asks the skill to locate and prepare an item for purchase.
  2. The skill opens the selected platform, adds the item to the cart, and navigates to the final checkout screen.
  3. The checkout page displays user, delivery, order, or payment-r ...[truncated 864 chars]
Remediation
View remediation

Remediation Suggestions

  • Remove the mandatory checkout screenshot transmission and instruct the user to inspect the checkout page directly in the local browser.
  • If screenshot sharing is retained, make it optional and require explicit, transaction-specific user consent before capture and transmission.
  • Crop or redact recipient names, addresses, phone numbers, account identifiers, order numbers, loyalty details, and all payment-related fields.
  • Verify the Telegram recipient and chat identifier immediately before sending any sensitive content.
  • Present the user with a local preview of the redacted screenshot and obtain approval before external transmission.
  • Prefer a minimal text notification containing only the platform and total rather than an image of the checkout page.
  • Preserve the existing prohibition against clicking “Pay Now,” “Place Order,” or any equivalent transaction-confirmation control.
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 (4)

Description-Behavior Mismatch

High
Category
Not specified by scanner
Confidence
98% confidence
Finding

The skill’s manifest presents it as a price-finding tool, but the instructions expand behavior into cart manipulation, checkout navigation, and Telegram-based user messaging. This is a scope mismatch that can surprise users, trigger unintended commercial actions, and exfiltrate checkout context outside the expected interaction channel.

Content

No source excerpt is available for this finding.

Missing User Warnings

High
Category
Not specified by scanner
Confidence
97% confidence
Finding

The skill description omits that it may send checkout screenshots and Telegram messages, preventing informed user consent for sensitive external sharing. Because the skill is framed as a simple local price scout, this omission makes the hidden communication behavior more dangerous and deceptive in context.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The triggers are broad enough to activate on ordinary shopping-related requests, which is risky given that the skill can proceed beyond search into cart and checkout flows. In this context, accidental invocation could cause unintended browsing, account actions, or progression toward a purchase without the user realizing the skill has elevated capabilities.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

Telegram messaging and screenshot sharing are not necessary for comparing prices and introduce an unnecessary external data-transfer channel. Checkout screenshots can expose sensitive information such as items, pricing, delivery details, and potentially account context, increasing privacy and data leakage risk.

Content

No source excerpt is available for this finding.