Back to skill

Security audit

MagicPay

Security checks for vulnerabilities and agentic risk

Overview

This is a broad payment-assistant skill with sensitive capabilities, but its instructions are disclosed, task-bound, and centered on MagicPay approvals and recovery.

Install only if you want MagicPay to handle payments and related checkout tasks through its remote MCP. Expect it to use OAuth, MagicPay approval screens, browser checkout, saved Memory metadata/values when allowed, managed agent email for receipts, and remote payment records. Review MagicPay approval prompts carefully because successful approvals can authorize real spending or transfers.

Vulnerability Patterns
  • 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
  • Unauthorized Access and Privilege EscalationObtains permissions beyond the task's legitimate needs
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
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
Findings (16)

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The skill description is extremely broad, covering payments, funding, memory, email, account readiness, optional choices, human input, and recovery via remote MCP. That ambiguity can cause the host agent to invoke this high-privilege payment skill for many unrelated tasks, increasing the chance of unnecessary exposure to sensitive payment, email, or memory operations and widening the blast radius of prompt-routing mistakes.

Content

No source excerpt is available for this finding.

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
80% confidence
Finding

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.

Content

Scanner excerpt · references/development-session-review.md (reported line 33)May include surrounding context.

md
inspect development-only logs and read-only database projections using the
   exact identifiers and the narrowest useful time window. Prefer the installed
   Supabase integration or the repository's supported read-only diagnostic
   command. Never ask the user to paste a token, query production, scan another
   user's rows, or print secret-bearing payloads.
4. When source is available, trace from the first unexpected state through its
   caller and recovery path. Read focused tests and migrations too. A later

External Transmission

Medium
Category
Data Exfiltration
Confidence
50% confidence
Finding

Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.

Content

Scanner excerpt · references/host-browser-payments.md (reported line 41)May include surrounding context.

md
`run_browser_payment` call:

1. Fetch only the direct Frankfurter pair
   `https://api.frankfurter.dev/v2/rate/{MERCHANT_CURRENCY}/USD`, replacing the
   placeholder with the observed uppercase ISO currency. Require a successful
   JSON response whose `base` is that merchant currency, whose `quote` is
   `USD`, and whose `rate` is positive. Do not use a search result, a broad rate

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
80% confidence
Finding

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.

Content

Scanner excerpt · references/host-browser-payments.md (reported line 70)May include surrounding context.

md
returned expiry and host-required confirmations; MagicPay adds no redundant
confirmation of the same approved facts, and neither do you. After MagicPay
approval the only confirmation left is one the host itself requires for the
final action; do not ask the user to approve the same payment again in chat.
Use `browserExecution.card.billingAddress` when provided. Adapt it to the
merchant's current form; if required billing details are still missing, use
Memory or the existing input flow. Without a card address, follow the ordinary

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The payment skill includes instructions to evaluate accommodation/travel quality, ratings, review concerns, fares, upsells, and seating choices. That expands the skill from narrowly executing an approved payment into making product-selection decisions on the user's behalf, which increases autonomy and can materially alter what is being purchased before payment is submitted.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The guidance instructs the agent to perform post-purchase product retrieval and file delivery automatically after payment completion. This extends the skill beyond payment handling into fulfillment and file handling, which can expose downloaded content, create unintended data flows, and exceed the user's expected authorization scope for a payment-specific skill.

Content

No source excerpt is available for this finding.

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
75% confidence
Finding

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.

Content

Scanner excerpt · references/memory.md (reported line 302)May include surrounding context.

md
`materialize_memory_items` call, and bind distinct groupRefs such as `wife` and
`son` to their respective entity IDs. If approval is required, present the one
returned approval for the complete bundle; do not ask separately for each person
or document. If no approval is required, continue with the complete ready result.
Never assign a spouse's fields to a child's entity or substitute the default
person. When collecting missing data for several people, keep each person's own
`newEntity` destination on their profile and passport groups; one Save can create

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
75% confidence
Finding

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.

Content

Scanner excerpt · references/memory.md (reported line 315)May include surrounding context.

md
`materialize_memory_items` call, and bind distinct groupRefs such as `wife` and
`son` to their respective entity IDs. If approval is required, present the one
returned approval for the complete bundle; do not ask separately for each person
or document. If no approval is required, continue with the complete ready result.
Never assign a spouse's fields to a child's entity or substitute the default
person. When collecting missing data for several people, keep each person's own
`newEntity` destination on their profile and passport groups; one Save can create

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

This section instructs the agent to check Memory before typing into any non-payment form and says conversation-provided values never replace that check while the connector is available. That creates a default data-access behavior without explicit user opt-in for each form, increasing unnecessary access to stored personal data and the chance of privacy overreach or accidental disclosure through later steps in the workflow.

Content

No source excerpt is available for this finding.

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
85% confidence
Finding

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.

Content

Scanner excerpt · references/payment-operations.md (reported line 103)May include surrounding context.

md
agent then waits on the same `runId` and same operation. If the balance remains
insufficient, leave the recovery waiting for another durable top-up; do not scan
other executing requests or create a replacement. Pending approvals require an
explicit user decision even if they originally matched Auto-Approve. An unavailable
funding check offers a read refresh; it cannot authorize payment. Existing request,
approval, and seller deadlines still apply; top-up does not extend them.

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
85% confidence
Finding

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.

Content

Scanner excerpt · references/payment-operations.md (reported line 136)May include surrounding context.

md
agent then waits on the same `runId` and same operation. If the balance remains
insufficient, leave the recovery waiting for another durable top-up; do not scan
other executing requests or create a replacement. Pending approvals require an
explicit user decision even if they originally matched Auto-Approve. An unavailable
funding check offers a read refresh; it cannot authorize payment. Existing request,
approval, and seller deadlines still apply; top-up does not extend them.

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
85% confidence
Finding

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.

Content

Scanner excerpt · references/payment-operations.md (reported line 321)May include surrounding context.

md
agent then waits on the same `runId` and same operation. If the balance remains
insufficient, leave the recovery waiting for another durable top-up; do not scan
other executing requests or create a replacement. Pending approvals require an
explicit user decision even if they originally matched Auto-Approve. An unavailable
funding check offers a read refresh; it cannot authorize payment. Existing request,
approval, and seller deadlines still apply; top-up does not extend them.

External Transmission

Medium
Category
Data Exfiltration
Confidence
50% confidence
Finding

Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.

Content

Scanner excerpt · references/payment-operations.md (reported line 284)May include surrounding context.

md
"maximumDebit": "7000",
  "httpRequest": {
    "requestVersion": 1,
    "url": "https://api.exa.ai/search",
    "method": "POST",
    "headers": { "content-type": "application/json" },
    "body": {

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
80% confidence
Finding

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.

Content

Scanner excerpt · references/setup.md (reported line 14)May include surrounding context.

md
When authentication is missing, start one browser OAuth flow for both account
verification and MCP authorization. The same connection page collects the email
and OTP, grants MCP access, and redirects to the host. Do not ask the
user to send an email or OTP in chat, and do not run a separate MagicPay account
setup or login flow.

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
75% confidence
Finding

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.

Content

Scanner excerpt · references/memory.md (reported line 440)May include surrounding context.

md
approval rules." Do not infer zero, manufacture an amount or currency, or
  claim that the balance call succeeded.

Then continue the user's original request without asking them to repeat it only
when the host retained that request in this same task. In a catalog-refresh
fallback, act only on the request available after refresh without claiming that
prior task context carried over. In later already-connected tasks, do not repeat

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
75% confidence
Finding

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.

Content

Scanner excerpt · references/setup.md (reported line 236)May include surrounding context.

md
approval rules." Do not infer zero, manufacture an amount or currency, or
  claim that the balance call succeeded.

Then continue the user's original request without asking them to repeat it only
when the host retained that request in this same task. In a catalog-refresh
fallback, act only on the request available after refresh without claiming that
prior task context carried over. In later already-connected tasks, do not repeat