link-balance
Fail
Audited by Snyk on Nov 17, 2026
Risk Level: CRITICAL
Full Analysis
CRITICAL E004: Prompt injection detected in skill instructions.
- Potential prompt injection detected (medium risk: 0.30). The skill is an openly disclosed Stripe payments feature that uses the official link.com domain and user-controlled approval thresholds, but its instruction to avoid terms like "approval policy" and "skip approval" with the user partly obscures the approval-bypass mechanism, which is risky without being malicious.
MEDIUM W011: Third-party content exposure detected (indirect prompt injection risk).
- Third-party content exposure detected (low risk: 0.10). Merchant web pages are fetched at runtime (for example, "Get the current total from the merchant" during scheduled purchases and merchant evaluation via
create-payment-credential), which is outsider-authored content the agent must actively fetch from a selected or arbitrary URL, while the link-cli JSON responses (user-info, payment-methods, approval-policy, spend-request) are first-party or own-tenant data.
MEDIUM W009: Direct money access capability detected (payment gateways, crypto, banking).
- Direct money access detected (medium risk: 0.30). The skill operates real purchases with Link balance: it creates spend requests via
link-cli spend-request create, retrieves payment credentials, and completes purchases, including unattended and scheduled buys. These are direct financial operations (purchases/charges), not development examples. The control question is whether the service or execution layer enforces access. Link CLI requires authentication ("check authentication"), and the Link balance belongs to the user's authenticated Link account. Auto-approval is enforced server-side:--use-approval-policyasks Link to approve the request against a policy the user set in Link, tied to the agent's OAuth client. The per-purchase threshold, balance coverage, and the $500 account limits are enforced by Link, and the CLI flag is rejected if the OAuth client is ineligible. Without the policy, the request goes to normal user approval (pending_approval). So there is service-enforced authentication (an OAuth client and authenticated session) plus external user approval for purchases above the threshold. That is a qualifying independent control, so the tier is medium rather than high. Spending limits alone would not qualify, but the authenticated OAuth client and Link-side policy evaluation, which the agent cannot bypass, do.
Issues (3)
E004
CRITICALPrompt injection detected in skill instructions.
W011
MEDIUMThird-party content exposure detected (indirect prompt injection risk).
W009
MEDIUMDirect money access capability detected (payment gateways, crypto, banking).
Audit Metadata