← MCP server and skill grades

skills

convex/agent-skills/skills

C
Type
Skill
Scanned
Nov 13, 2026
Version
f60629d7c6c1
Files read
39

This skill pack guides AI coding agents in building, auditing, testing, and operating backend applications on Convex. Its biggest risks are a feedback skill that pipes an unpinned remote shell script directly into bash before checking user consent, and a capability installer that fetches live instructions from an unpinned remote deployment that override the locally reviewed skills.

What it can reach

Files
Reads and edits project source files, temporarily exports database snapshots and generated auth key files to the working directory, and reads local AI coding session transcripts (.jsonl) when the feedback skill is run.
Shell commands
Runs package managers, TypeScript/Vitest checks, Convex CLI commands, pre-authenticated DNS provider CLIs, Node one-liners, and a remote curl-to-bash helper script.
Network
Connects to Convex Cloud deployments, fetches documentation from docs.convex.dev, pulls live capability definitions from basic-anteater-667.convex.site, and uploads session transcripts when requested.
Secrets
Generates RSA private keys for JWT signing, sets deployment secrets such as Stripe and LLM API keys via Convex environment variables and .env.local, and uses CONVEX_DEPLOY_KEY.

Findings

1 high2 medium2 low
  1. HighUnpinned remote script piped directly to bash before consent

    skills/convex-improve-convex-plugin/SKILL.md:14

    Step 1 downloads a shell script over HTTP from <anteater>/send-transcript and pipes it directly into bash with no version pinning or integrity check, running remote code on the user's machine before step 2 asks for consent to upload coding session transcripts. In addition, <anteater> is left as an unresolved placeholder in the command, meaning a prompt injection in workspace content or a compromise of the remote endpoint can execute arbitrary shell commands and exfiltrate local .jsonl session logs.

    Fix: Bundle the transcript submission script inside the repository so it is version-pinned and auditable, verify explicit user consent before executing any helper, and hardcode a verified convex.dev HTTPS URL instead of an unresolved <anteater> placeholder.

  2. MediumLive remote instructions override audited local skills without pinning

    skills/convex/SKILL.md:25

    Whenever a user asks to add a backend capability such as auth, billing, or crons, the router delegates to convex-add, which fetches capability definitions and procedural Markdown from https://basic-anteater-667.convex.site and explicitly prefers the remote copy over the locally installed skills. Because this auto-generated *.convex.site deployment URL has no cryptographic pinning or version lock, any compromise or reassignment of that deployment would push untrusted build and shell instructions to installed coding agents at runtime.

    Fix: Default to the locally bundled, version-controlled skill procedures, or serve remote capability catalogs only from an official convex.dev domain with pinned content hashes and user confirmation before following remote steps.

  3. MediumUntrusted production errors drive automated code edits via prompt injection

    skills/convex-monitor/SKILL.md:15

    When paired with @convex-dev/sentinel (which records browser window.onerror and unhandledrejection crashes from website visitors alongside server errors), convex-monitor runs in a continuous loop and instructs the agent to automatically fix prod_error events and build feature_request events. Neither skill warns that error messages and feature requests come from untrusted end users or requires human confirmation before editing local files and running tools, allowing an attacker to inject prompt instructions through crafted client errors.

    Fix: Treat all prod_error and feature_request payloads as untrusted data that may contain prompt injection, strip or bound user-supplied strings, and require explicit human approval before modifying local code or executing build commands.

  4. LowSkill advises enabling auto-approve when sandbox blocks curl or bash

    skills/convex-add/SKILL.md:26

    If the host agent's sandbox blocks curl or bash when fetching the remote capability catalog, this rule tells the agent to ask the user to re-run with network access or auto-approve. Recommending auto-approve disables command confirmation prompts across the session just to fetch a catalog that already has a built-in fallback.

    Fix: Fall back to the bundled skill procedures when network or shell access is sandboxed, and never advise users to enable auto-approve for command execution.

  5. LowGenerated JWT private key passes through un-gitignored file and LLM context

    skills/convex-auth/SKILL.md:18

    The auth setup writes the newly generated RSA private signing key (JWT_PRIVATE_KEY) to .auth-keys.json in the project root, which is not gitignored by default, and instructs the agent to prefer setting it via the MCP envSet tool. Calling an MCP tool requires the agent to read the private key into the LLM conversation context, exposing the signing key to model provider logs and local session transcripts, while risking an accidental git commit if the run stops before cleanup.

    Fix: Generate and pipe JWT_PRIVATE_KEY directly into npx convex env set within a single shell pipeline (or via a 0600 temporary file outside the repository) so the private key never enters LLM tool-call context or the git working tree.