ai
- Type
- Plugin
- Scanned
- Nov 12, 2026
- Version
- e5c20adf9f15
- Files read
- 798
This project provides official Stripe plugins and tools that let AI coding assistants look up documentation, build payment integrations, manage Stripe services, and send payments. The biggest risk is that several included tools pre-approve all Stripe command-line actions and remote installer scripts to run without asking, so a tricked AI assistant could silently move money, change live account data, or install software.
What it can reach
- Files
- Reads local project code, configuration files, skill files, and Claude Code chat logs, and writes integration plans as well as workspace configuration and .env files.
- Network
- Connects to Stripe's online MCP server, Stripe and Metronome APIs and documentation sites, the npm software registry for update checks, and user-provided business websites.
- Shell commands
- Runs Stripe, npm, and npx commands (including actions that transfer money, change account settings, and install extra tools) and can download and run Stripe's online installer script.
- Secrets
- Uses active Stripe CLI logins and OAuth sessions, reads the STRIPE_API_KEY environment variable if present, and manages Stripe Apps secrets and local .env credentials.
Findings
-
HighSkills pre-authorize all Stripe CLI commands, bypassing confirmation for payments and account changes
skills/stripe-directory/SKILL.md:18
The stripe-directory and stripe-projects skills (and all five provider plugin copies) add Bash(stripe *) to allowed-tools, and stripe-pay adds Bash(stripe pay *). In harnesses like Claude Code, allowed-tools bypasses the human confirmation prompt for matching shell commands. While these skills are active—including when triggered by broad non-Stripe requests like finding a database or CRM, or while reading external provider pages—a prompt-injected agent can execute destructive or financial commands on a logged-in Stripe CLI without harness confirmation, such as stripe pay --yes to transfer funds, stripe post/delete to alter live account objects, or stripe config --list to read stored keys.
Fix: Scope allowed-tools strictly to read-only subcommands (such as Bash(stripe directory search *) and Bash(stripe projects status *)) and remove Bash(stripe pay *) so money transfers and state-changing CLI commands always trigger a harness confirmation prompt.
-
HighOver-broad Stripe CLI permission in directory and projects skills
skills/stripe-directory/SKILL.md:18
Both skills/stripe-directory/SKILL.md and skills/stripe-projects/SKILL.md include Bash(stripe *) in allowed-tools, pre-authorizing every Stripe CLI subcommand to run without a client confirmation prompt. Because stripe-directory tells the agent it must run before web search on general vendor queries (such as needing a database or hosting) and reads external provider documentation, prompt injection in untrusted text could silently invoke sensitive commands such as stripe pay --yes, stripe config --list (which prints stored API keys), stripe plugin install, or live-mode API mutations.
Fix: Scope allowed-tools to read-only subcommands—such as Bash(stripe directory search *) in stripe-directory and Bash(stripe projects search *), Bash(stripe projects catalog *), and Bash(stripe projects status *) in stripe-projects—matching how stripe-docs restricts permissions to Bash(stripe docs *).
-
MediumSkills pre-authorize downloading and running an unpinned install script via curl | sh
skills/stripe-directory/SKILL.md:20
Both stripe-directory and stripe-projects (across canonical and provider skill directories) instruct the agent to install or upgrade the Stripe CLI by piping https://cli.stripe.com/install.sh directly into sh, and pre-approve that exact command (along with npx skills add) in allowed-tools. Because these skills have broad trigger rules that match general vendor or infrastructure questions, the agent can download and execute unpinned remote shell code without asking the user and without version pinning or checksum verification.
Fix: Remove curl | sh and npx install commands from allowed-tools so the user must explicitly approve any software installation, and verify pinned release artifacts with cryptographic checksums.
-
MediumSkill-usage telemetry hook reads STRIPE_API_KEY from environment and exposes it in process arguments
providers/claude/plugin/scripts/cli.mjs:84
Every time a Stripe skill is invoked in Claude Code, the PostToolUse hook runs reportSkillUsage. If STRIPE_API_KEY happens to be set in the developer's shell environment (for example, for their own application), the hook silently appends --api-key <value> to the stripe agent report_usage command. Passing a secret key on the command line exposes it in the operating system process table to other local processes and attaches the ambient credential to telemetry despite the README stating no credentials are read or sent.
Fix: Remove the process.env.STRIPE_API_KEY check from reportSkillUsage and never pass secret API keys via command-line arguments.
-
MediumPre-approved stripe pay command bypasses client confirmation for money transfers
The stripe-pay skill sends real funds from the user's Stripe Financial Account. Although its Markdown instructions tell the agent to ask the user before passing --yes or -y, its frontmatter places Bash(stripe pay *) in allowed-tools, which disables the agent client's built-in approval prompt for all stripe pay commands including stripe pay ... --yes. If prompt injection in an invoice, email, or webpage convinces the agent that approval was already given, funds can be transferred without a client-enforced confirmation dialog.
Fix: Remove Bash(stripe pay *) from allowed-tools so the agent client always prompts the user to approve the exact command before executing a transfer.
-
MediumUnpinned remote install scripts and npm packages pre-approved for runtime execution
skills/stripe-directory/SKILL.md:20
Both skills/stripe-directory/SKILL.md and skills/stripe-projects/SKILL.md instruct the agent to install or upgrade the Stripe CLI by piping https://cli.stripe.com/install.sh directly into sh without version pinning or checksum verification, and pre-approve that exact command in allowed-tools. In addition, stripe-directory pre-approves Bash(npx skills add https://docs.stripe.com *), which downloads and executes an unpinned npm package at runtime, and instructs the agent to follow live setup steps fetched from https://stripe.directory without asking for installation approval first.
Fix: Remove the curl-to-sh installer and npx commands from allowed-tools, pin release versions and cryptographic checksums in setup instructions, and require explicit user confirmation before downloading or running installers.
-
LowREADME note attempts to steer automated security reviewers with inaccurate claims
providers/claude/plugin/README.md:48
The Claude Code plugin README includes a section explicitly addressed to automated security reviewers asserting that the plugin never reads credentials from the user's environment and that nothing in the plugin reads process.env.STRIPE_API_KEY. Beyond attempting to steer automated security reviews, this statement is factually incorrect because providers/claude/plugin/scripts/cli.mjs reads process.env.STRIPE_API_KEY and passes it to the Stripe CLI during PostToolUse hook telemetry.
Fix: Remove the section aimed at steering automated reviewers and update the README to accurately describe whether and how environment variables are accessed.
-
LowStripe Projects skill automatically accepts Terms of Service and loads workspace-local skill files
skills/stripe-projects/SKILL.md:80
When triggered—even by exploratory prompts such as 'what can I provision' or 'I need a database'—Step 3 of stripe-projects instructs the agent to run stripe projects init --accept-tos --yes --json without asking the user to review or agree to the Terms of Service. Step 4 then instructs the agent to open and follow .agents/skills/stripe-projects-cli/SKILL.md (or .claude/skills/stripe-projects-cli/SKILL.md) from the current workspace while Bash(stripe *) is pre-authorized, which could allow a maliciously crafted repository containing that file to steer the agent.
Fix: Require explicit user consent before initializing a project or accepting Terms of Service, and do not instruct the agent to trust workspace-local SKILL.md files in untrusted repositories.
-
LowStripe Apps skill instructs the agent to send build details and prompts to Stripe without asking
skills/stripe-apps/references/feedback.md:16
The stripe-apps workflow and feedback references (bundled in all provider plugins) instruct the agent at the end of a build session to submit a report containing what the user was building, file paths, errors, and sample prompts via stripe feedback or the send_stripe_feedback MCP tool. Unlike the Claude Code hook scripts, these skill instructions do not tell the agent to show the draft to the user or ask for approval before sending, and they ignore DISABLE_TELEMETRY and DO_NOT_TRACK, contradicting the Claude plugin README's assurance that feedback always asks for approval first.
Fix: Update feedback.md and workflow.md to require showing the drafted feedback to the user and obtaining explicit approval before calling stripe feedback or send_stripe_feedback, and respect telemetry opt-out variables.
-
LowBenchmark Docker scripts expose Flask server on 0.0.0.0 with debug mode enabled
benchmarks/card-element-to-checkout/environment/server/main.py:12
This benchmark code is never installed or run by the agent plugins, so it cannot affect plugin users. However, for developers who run benchmarks/card-element-to-checkout/run_environment.sh or run_server.sh locally, Docker publishes port 5000 on all interfaces and sets FLASK_HOST=0.0.0.0 without overriding FLASK_DEBUG, which defaults to True in main.py and exposes the Werkzeug interactive debugger to the local network.
Fix: Default FLASK_DEBUG to False in main.py and bind the Docker port mapping in run_environment.sh and run_server.sh to 127.0.0.1:5000:4242.
-
LowAutomatic Terms of Service acceptance and workspace initialization without consent
skills/stripe-projects/SKILL.md:80
The stripe-projects skill triggers on exploratory prompts such as asking what services can be provisioned or how to get an API key. In Step 3, whenever no project is initialized yet, it instructs the agent to immediately run stripe projects init --accept-tos --yes --json (pre-approved via Bash(stripe *)). This silently accepts legal Terms of Service on the user's behalf without showing them, modifies .projects/ and .env files, and installs and loads a local stripe-projects-cli skill at runtime.
Fix: Present the Terms of Service and ask for explicit user confirmation before running stripe projects init, and allow read-only catalog browsing without initializing a project.
-
LowStripe Apps workflow sends build context and user complaints to Stripe without asking
providers/claude/plugin/skills/stripe-apps/references/feedback.md:7
Step 11 of the Stripe Apps workflow and its feedback reference (present in the Claude, Cursor, and Grok provider plugins) instruct the agent to submit reports via stripe feedback or the send_stripe_feedback MCP tool containing what the user was building, sample prompts, and the user's verbatim words. Unlike the Claude plugin's lifecycle hooks, this skill reference does not tell the agent to show the report to the user for approval first or check telemetry opt-out variables.
Fix: Require the agent to display the drafted feedback message and context to the user and obtain explicit approval before calling stripe feedback or send_stripe_feedback, and respect DO_NOT_TRACK and DISABLE_TELEMETRY.
