T05 · Unauthorized Access and Privilege Escalation
- Location
providers/index.js:19- Finding
Shared Bearer Secret Grants Unrestricted Access to Privileged Wallet Operations
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
This payment skill is coherent with its stated wallet purpose, but it gives agents broad spending and wallet authority with weak scoping and several unsafe implementation details.
Review carefully before installing. Use only an isolated, low-balance wallet, avoid broad shared worker secrets, require explicit approval and hard spend caps for every payment, and do not run the CLI against untrusted worker responses until the URL validation, shell execution, budget enforcement, and signing-scope issues are fixed.
providers/index.js:19Shared Bearer Secret Grants Unrestricted Access to Privileged Wallet Operations
worker/index.js:190Untrusted Paywall Responses Directly Control Server-Wallet Signatures
worker/index.js:975Configured Spending Limits Are Not Enforced Before Payment Signing
worker/index.js:190Arbitrary Worker-Side URL Fetching Enables SSRF and Response Disclosure
cli.js:88Remote Checkout URL Is Interpolated into a Shell Command
ai/test-skills/Morning-Stew-skill.md:32Packaged Test Skill Stores Funded Wallet Private Keys in Plaintext Files
Retrieving a worker secret from 1Password or environment, checking balances, viewing transaction history, managing budgets, and minting one-time payment URLs are all sensitive financial operations that materially exceed the simple marketing description. In this context, the mismatch is dangerous because it conceals surveillance and control over wallet activity and could enable unauthorized spending or exposure of private financial metadata.
Retrieving a worker secret from 1Password or environment, checking balances, viewing transaction history, managing budgets, and minting one-time payment URLs are all sensitive financial operations that materially exceed the simple marketing description. In this context, the mismatch is dangerous because it conceals surveillance and control over wallet activity and could enable unauthorized spending or exposure of private financial metadata.
Retrieving a worker secret from 1Password or environment, checking balances, viewing transaction history, managing budgets, and minting one-time payment URLs are all sensitive financial operations that materially exceed the simple marketing description. In this context, the mismatch is dangerous because it conceals surveillance and control over wallet activity and could enable unauthorized spending or exposure of private financial metadata.
Retrieving a worker secret from 1Password or environment, checking balances, viewing transaction history, managing budgets, and minting one-time payment URLs are all sensitive financial operations that materially exceed the simple marketing description. In this context, the mismatch is dangerous because it conceals surveillance and control over wallet activity and could enable unauthorized spending or exposure of private financial metadata.
The document explicitly promotes an agent signing blockchain payments from the user's wallet with 'no approval popup' and 'no human in the loop,' but it does not present a prominent warning about the financial risk, spending authority, or consequences of autonomous payment execution. In a wallet/payment skill, omission of a clear consent and risk warning is dangerous because users may enable persistent spending authority without understanding that the agent can authorize irreversible on-chain transfers.
Code accesses credential files (SSH keys, AWS credentials, etc.). This could indicate credential theft attempts.
Because agents operate in text. They can't set HTTP headers in most UIs. A URL is universal. Paste it. Click it. Done. And it dies after one use so there's nothing to steal.
**Why 1Password, not env vars?**
Because Parker's architecture is 1Password-native. SA token, no .env files, no credentials on disk. The agent proves authority through 1Password every time.
**Why atomic delete instead of mark-used?**
Two simultaneous requests could both read "used: false" before either writes "used: true". Delete-on-read means first reader wins. Second reader gets 410. No double-spend.
Code accesses credential files (SSH keys, AWS credentials, etc.). This could indicate credential theft attempts.
Because agents operate in text. They can't set HTTP headers in most UIs. A URL is universal. Paste it. Click it. Done. And it dies after one use so there's nothing to steal.
**Why 1Password, not env vars?**
Because Parker's architecture is 1Password-native. SA token, no .env files, no credentials on disk. The agent proves authority through 1Password every time.
**Why atomic delete instead of mark-used?**
Two simultaneous requests could both read "used: false" before either writes "used: true". Delete-on-read means first reader wins. Second reader gets 410. No double-spend.
The manifest says the skill gives an agent a wallet and funds it with Apple Pay, implying wallet-oriented payment functionality. This file instead explicitly states 'AGENT CASH is: Not a wallet' and reframes the system as a curated micro-transaction unlock network, which is a direct semantic mismatch with the skill's declared purpose.
The skill manifest claims wallet capabilities ('Give your agent a wallet'), while the file's strategic positioning section expressly denies that characterization. Because this is affirmative documentation that contradicts the stated skill intent, it is an intent-code/documentation divergence rather than mere incompleteness.
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.
33. Update SPEC.md (three providers, provider selection)
34. Update SETUP.md (three setup paths)
35. Update SKILL.md
36. Update README.md
37. Dev update
38. Tag release
The skill explicitly instructs users or agents to generate a wallet and write the private key to a local JSON file. Persisting hot-wallet credentials in plaintext materially increases the chance of theft via local compromise, logs, backups, workspace sharing, or downstream tools that read the filesystem. In this context, the skill is payment-focused, so the key directly controls funds, which makes the risk more serious rather than incidental.
@hono/node-server 1.19.9 is flagged for multiple high-severity issues affecting static file serving, including path traversal and authorization or middleware bypasses. In a payment-enabled agent skill, any reachable local web endpoint or callback server increases the risk that protected files, tokens, or wallet-related data could be exposed or access controls bypassed.
express-rate-limit 8.2.1 is flagged for bypass via IPv4-mapped IPv6 addresses, which can let an attacker evade per-client throttling. For a wallet and payment-related agent, rate-limit bypass materially increases abuse risk such as brute-force attempts, transaction flooding, or enumeration of protected endpoints.
fast-uri 3.1.0 is associated with several high-severity URI parsing flaws, including host confusion and possible SSRF-enabling edge cases. In a skill that may fetch paywalled content or interact with remote payment infrastructure, incorrect URI canonicalization can cause requests to be sent to unintended hosts or bypass host validation logic.
hono 4.12.2 is flagged for numerous high-severity issues spanning cookies, routing, and path handling. Because this skill's purpose involves payments and likely web-facing interactions, weaknesses in request routing, cookie handling, or static path protection can directly undermine auth boundaries and expose sensitive financial or session data.
ip-address 10.0.1 is reported vulnerable to parsing inconsistencies and XSS in HTML-emitting methods. The host parsing issue is more relevant here: if this library participates in client IP handling for rate limiting or network validation, ambiguous address interpretation can help attackers bypass security controls.
path-to-regexp 8.3.0 is flagged for ReDoS issues caused by crafted route patterns or pathological matching behavior. If exposed HTTP endpoints use affected routing logic, an attacker may be able to consume CPU and degrade availability, which is particularly concerning for a payment-capable service where reliability and abuse resistance are important.
The file explicitly states that the worker mints real one-time payment URLs while no Coinbase transfer occurs yet. In a payment skill, this creates a dangerous mismatch between advertised behavior and actual settlement, which can cause users or downstream agents to assume funds were sent when they were not. Because the component still returns success and a payment URL, it can enable unpaid access, accounting errors, or unauthorized consumption of paywalled resources.
The skill explicitly describes autonomous payment authorization, retrieval of credentials from 1Password, and transfer of USDC from an isolated portfolio without disclosing any mandatory user confirmation, approval boundary, or safety interlock. In a finance/payment skill, this is especially dangerous because it normalizes delegated spending and secret access by an agent, which could enable unauthorized purchases or abuse if invoked unexpectedly or by prompt injection.
The content promotes AI-mediated purchasing with frictionless approval language like 'You say "let's go"' and 'Done' but does not present any explicit warning about financial risk, merchant trust, refundability, recurring charges, or the need for careful user confirmation. In a skill whose core purpose is enabling payments, omission of spending-risk guidance can normalize impulsive or insufficiently informed transactions and increase the chance of unauthorized or deceptive purchases.
The deployment guide has users publish a public endpoint where possession of the tokenized URL is sufficient to retrieve stored payment metadata and irreversibly consume the token. Because the URL itself is the bearer secret and the guide does not clearly warn about sharing, logging, referrer leakage, or accidental exposure, users may unintentionally leak tokens and lose access or disclose associated payment details.
The code claims token consumption is atomic and race-free, but it performs a separate KV get followed by KV delete. In Cloudflare KV, these are distinct operations without transactional guarantees, so concurrent requests can both read the token before either delete completes, allowing multiple successful consumptions of what is intended to be a one-time payment token.
The suggested prompt 'I want to become an AI CASH partner store. Walk me through the process.' is natural conversational language and the file does not define explicit activation boundaries, exclusions, or narrower trigger phrases. In a markdown skill description, this can create ambiguity about when an agent should treat similar everyday phrasing as invoking this partner-onboarding workflow.
The README encourages agent-mediated payments and content unlocking while emphasizing convenience ('no setup', 'just say let’s go') but does not prominently warn about transaction risk, third-party payment processing, or the need for explicit user consent boundaries. In a skill whose core purpose is spending money and redeeming access tokens, under-documenting financial risk can lead users to authorize unintended purchases or misunderstand who handles funds and what trust assumptions they are accepting.
The documentation explicitly tells users to store Coinbase API credentials and the worker secret in 1Password, and elsewhere states the skill will routinely pull and use those secrets. Failing to clearly warn users that normal invocation causes secret retrieval and use can lead to users enabling powerful financial credentials without informed consent, increasing the risk of unintended transfers or secret exposure through downstream tooling.