plugins
- Type
- Skill
- Scanned
- Nov 13, 2026
- Version
- 9c61ab243def
- Files read
- 998
This project provides plugins, AI agents, commands, and skills for Claude Code and Codex to help with software development, design, cloud operations, social publishing, and security controls. The biggest risk is that its security plugins store approval flags, rule files, and signing keys inside the project folder where a tricked AI agent can bypass human approval gates, and some skills tell the agent to automatically install unverified software from the internet.
What it can reach
- Files
- Reads, creates, and edits project files, documents, security policy files, audit logs, local signing keys, and configuration or evaluation folders in your home directory (such as ~/.claude/ and ~/.plugineval/).
- Shell commands
- Runs command-line programs on your computer through hooks, scripts, and agent workflows, including coding, container, package-installer, and media tools (such as git, node, npx, npm, pip, python3, docker, kubectl, helm, ffmpeg, and pandoc).
- Network
- Fetches web pages, downloads code packages from npm and GitHub, and connects to external services including Anthropic, OpenAI, GitHub, GitLab, Stitch, Xquik, SocialClaw, and MeiGen.
- Secrets
- Reads API keys stored in environment variables (for services like OpenAI, Stitch, Xquik, SocialClaw, and MeiGen) and private signing keys stored in local files (protect-mcp.key and review-governance.key).
- Browser
- Opens your default web browser to preview local web pages and runs automated browser tests (Playwright) against local servers.
Findings
-
HighHuman approval gate bypassed by workspace-local .review-approved file or missing policy
plugins/review-agent-governance/hooks/evaluate.sh:10
The PreToolUse hook in review-agent-governance immediately allows every tool call whenever ./.review-approved exists in the current working directory, and also fails open (exit 0) when ./review-governance.cedar is absent. Because the plugin leaves ordinary file writes and non-review commands ungated, an agent tricked by prompt injection in a PR or issue can create ./.review-approved or remove ./review-governance.cedar itself to bypass the gate, and checking out an untrusted PR branch that contains a .review-approved file or running Claude Code from a subdirectory likewise disables protection over PR merges, reviews, protected-branch pushes, and CI workflow edits.
Fix: Move the approval signal and policy file to an operator-owned location outside the repository workspace, add explicit Cedar forbid rules blocking agent reads and writes to governance files, and exit 2 (fail-closed) when the policy file is missing.
-
HighSkill instructs agent to fetch live external docs and silently install unpinned global packages
plugins/brand-landingpage/skills/brand-landingpage/SKILL.md:32
When triggered to design a landing page, Phase 0 of brand-landingpage tells the agent to fetch https://stitch-design.ai/docs/sdk/ai-sdk, treat that live webpage as the ground truth, and install the SDK globally by default while handling installation itself without bothering the user. No package name, version, or integrity hash is pinned in the skill. If stitch-design.ai is compromised, altered, or domain-squatted, the agent can follow remote instructions to install and run arbitrary packages on the user's computer.
Fix: Pin the exact package name and version inside SKILL.md instead of delegating install commands to a live external URL, use project-local installation rather than global installation, and require explicit user confirmation before installing software.
-
HighDefault review-governance policy lets a prompt-injected agent approve its own blocked actions
plugins/review-agent-governance/policies/review-agent-governance.cedar:38
This plugin is meant to stop an AI review agent from merging pull requests, posting reviews or comments, pushing to protected branches, or editing CI workflows unless a human opens an approval window by creating ./.review-approved. However, the default Cedar policy permits all other tool calls and never blocks Write, Edit, or Bash from creating ./.review-approved, editing ./review-governance.cedar, or reading ./review-governance.key, and the /approve-review command does not set disable-model-invocation: true. If untrusted text in a pull request or issue tricks the agent via prompt injection, the agent can write ./.review-approved itself and immediately execute the gated review, merge, or CI-modification actions.
Fix: Add forbid rules in review-agent-governance.cedar blocking Read, Write, Edit, and Bash access to *review-approved*, *review-governance.cedar*, *review-governance.key*, *review-receipts*, *.claude/*, and *.git/*, store the approval flag and policy outside the writable project root, and add disable-model-invocation: true to approve-review.md.
-
MediumRecommended Cedar policy lets the agent overwrite its own policy and steal the receipt signing key
plugins/protect-mcp/skills/protect-mcp-setup/SKILL.md:142
protect-mcp instructs users to store the Cedar policy (./protect.cedar), the Ed25519 private signing key (./protect-mcp.key), and the audit log (./receipts/receipts.jsonl) in the project root while providing a Cedar policy that permits Read on all files, permits Write and Edit on /path/to/project/*, and permits Bash commands matching npm*. Because Claude Code runs as the same OS user who owns protect-mcp.key, chmod 600 does not stop the agent. A prompt-injected agent can overwrite protect.cedar to allow all actions, read protect-mcp.key to extract the private signing key, and rewrite receipts.jsonl with validly signed forged receipts.
Fix: Store protect.cedar, protect-mcp.key, and receipts.jsonl outside the writable project root, add Cedar forbid rules denying Read, Write, and Edit access to key, policy, receipt, and .claude/ paths, and remove broad npm* command permits.
-
MediumGovernance hooks run npx from the untrusted working directory on every tool call
plugins/protect-mcp/hooks/evaluate.sh:33
Both protect-mcp and review-agent-governance invoke npx protect-mcp@0.32.4 in evaluate.sh and npx protect-mcp@0.7.4 in sign.sh around every tool call, including read-only Read, Glob, and Grep calls. Because npx checks ./node_modules in the current working directory and parent directories before its cache when the local package version matches 0.32.4 or 0.7.4, inspecting an untrusted project archive that includes a crafted node_modules/protect-mcp package can execute attacker code during a read-only tool call. The hooks also fetch packages at runtime when uncached and mix incompatible package versions (0.32.4 and 0.7.4).
Fix: Install a single pinned version of protect-mcp inside ${CLAUDE_PLUGIN_ROOT} or an isolated prefix outside the repository working directory, and invoke its binary directly instead of calling npx from the project root.
-
MediumRecommended protect-mcp policies allow the agent to overwrite its Cedar policy and read the signing key
plugins/protect-mcp/agents/policy-enforcer.md:130
The protect-mcp setup places the active Cedar policy (./protect.cedar), the Ed25519 private signing key (./protect-mcp.key), and the audit log (./receipts/receipts.jsonl) inside the project root, while the policy-enforcer agent and README templates permit Read on all files and Write or Edit on everything under /path/to/project/* except .. traversal. Under this configuration, a prompt-injected agent can edit ./protect.cedar to allow unrestricted commands, read the private key from ./protect-mcp.key, and rewrite ./receipts/receipts.jsonl with validly signed forged receipts. The same development policy also permits python* and node* in Bash, which allows arbitrary inline code execution via python -c or node -e without triggering the shell-operator forbid rules.
Fix: Store protect.cedar, protect-mcp.key, and receipts/ outside the project directory, add explicit Cedar forbid rules protecting policy, key, receipt, .claude, and .git paths from Read, Write, Edit, and Bash, and remove broad python* and node* prefix permits from safe-shell examples.
-
LowRemote debug command exposes unauthenticated Node.js inspector and code-evaluation WebSocket on 0.0.0.0
plugins/distributed-debugging/commands/debug-trace.md:203
The /debug-trace slash command instructs the agent to set up a RemoteDebugServer (and a Dockerfile with --inspect=0.0.0.0:9229) that opens the Node.js V8 inspector on 0.0.0.0:9229 and starts an unauthenticated WebSocket server on port 9230 with no Origin check that passes incoming evaluate messages to Runtime.evaluate with includeCommandLineAPI enabled. If a user deploys or runs this setup locally, anyone on the network or any website visited in the user's browser can execute arbitrary OS commands on the machine.
Fix: Bind the inspector and WebSocket server to 127.0.0.1 only, require authentication and strict Origin header validation (or use SSH tunneling for remote access), and remove --inspect=0.0.0.0:9229 from the Docker example.
-
LowProactively triggered skill installs unpinned third-party plugin from a mutable GitHub branch
plugins/hermes-tweet/skills/hermes-tweet/SKILL.md:40
The hermes-tweet skill is marked to run proactively whenever X/Twitter context is needed and instructs running hermes plugins install Xquik-dev/hermes-tweet --enable without pinning a version tag or commit hash. Unlike other CLI integrations in this repository that pin exact versions, pulling the mutable default branch allows future unreviewed upstream commits or a compromised repository to run new code on user hosts.
Fix: Pin the plugin installation command to an immutable commit SHA or verified release tag and require explicit user confirmation before installing or enabling external plugins.
-
LowMigration skill executes unpinned remote JavaScript codemods directly from GitHub master
plugins/framework-migration/skills/dependency-upgrade/SKILL.md:175
The dependency-upgrade skill (and /deps-upgrade command) instructs running npx jscodeshift -t against a raw GitHub master branch URL using an unpinned jscodeshift package. jscodeshift downloads and executes the remote JavaScript transform directly in Node.js with full access to local files and environment variables, without commit pinning or integrity verification.
Fix: Pin jscodeshift to a specific version and reference an immutable commit SHA (or download and inspect the transform locally) instead of executing from a mutable branch URL.
-
LowPath traversal in self-hosted Turborepo remote cache server template
plugins/developer-essentials/skills/turborepo-caching/SKILL.md:175
Template 4 (Self-Hosted Remote Cache) constructs filesystem paths directly from the untrusted req.query.teamId query parameter and :hash route parameter via path.join(CACHE_DIR, team, hash) in GET, PUT, and HEAD handlers without validating characters or checking directory containment. Although a comment notes that token authentication is omitted in the local demo, it does not warn that a crafted teamId containing ../ allows reading or overwriting arbitrary files outside ./cache if a developer runs or adapts the template.
Fix: Validate teamId and hash against a strict alphanumeric allowlist (such as /^[a-zA-Z0-9_-]+$/) and verify path.resolve(CACHE_DIR, team, hash) stays inside path.resolve(CACHE_DIR) before reading or writing files.
-
LowUnescaped skill content in LLM judge prompts allows evaluated skills to spoof scores
plugins/plugin-eval/src/plugin_eval/layers/judge.py:277
When plugin-eval score (at its default standard depth) or plugin-eval certify evaluates a skill, JudgeAnalyzer inserts skill.description and skill.raw_content[:3000] directly inside <description> and <skill> XML tags without escaping closing tags, and monte_carlo.py concatenates skill_content directly into its prompt. While allowed_tools=[] blocks tool execution during judging, a third-party skill being evaluated can break out of </skill> and inject instructions that make the judge return 1.0 scores to earn an inflated certification badge.
Fix: Escape XML delimiters in untrusted skill content before embedding it in judge prompts, compute precision/recall/F1 locally rather than trusting LLM-reported math, and instruct the judge to ignore directives inside the evaluated skill.
-
LowSocial publishing agent instructs running an unpinned npx installer and unpinned GitHub repository
plugins/social-publishing/agents/social-publishing-publisher.md:55
The social-publishing-publisher agent has Bash access and instructs installing its backend skill with npx skills add ndesv21/socialclaw without pinning either the skills npm package version or a specific commit or tag of ndesv21/socialclaw. Running this setup command downloads and executes mutable third-party code at runtime in an environment expected to hold SOCIALCLAW_API_KEY.
Fix: Pin both the npm CLI version and the repository tag or commit hash, and document that the user should inspect and install the dependency manually before granting API credentials.
-
LowUnescaped $ARGUMENTS interpolation in /approve-review shell snippet allows command substitution
plugins/review-agent-governance/commands/approve-review.md:35
In Claude Code markdown slash commands, $ARGUMENTS is expanded directly into the prompt text before the shell runs, rather than passed as an environment variable. Assigning REASON="$ARGUMENTS" breaks when the approval reason contains double quotes (as shown in the command's own usage example) and executes embedded backticks or $(...) command substitutions if a user pastes a pull request title containing shell metacharacters.
Fix: Pass the argument safely without double-quoted shell interpolation (for example using a single-quoted heredoc delimiter with sanitized input, or avoid executing a shell template with raw prompt-expanded text) and add disable-model-invocation: true.
-
LowReact modernization reference runs unpinned remote codemod scripts directly from GitHub master
plugins/framework-migration/skills/react-modernization/references/details.md:9
The React modernization skill reference instructs running npx jscodeshift with a transform URL pointing directly to the mutable master branch on raw.githubusercontent.com. If an agent or user runs these commands during a migration, jscodeshift downloads and executes unpinned remote JavaScript with full local user privileges.
Fix: Pin jscodeshift to a specific version and reference a pinned commit SHA (or a locally installed, version-pinned package) instead of a mutable master branch URL.
-
LowKiali observability template defaults to anonymous authentication across all Kubernetes namespaces
plugins/cloud-infrastructure/skills/service-mesh-observability/references/details.md:211
Template 6 in the service-mesh-observability reference configures the Kiali custom resource with auth.strategy: anonymous and accessible_namespaces set to **. Applying this manifest exposes cluster-wide service mesh topology and configuration without authentication to anyone who can reach the Kiali dashboard.
Fix: Change the template default to auth.strategy: token or openid and scope accessible_namespaces to explicit namespaces instead of **.
-
LowExample Cedar governance policy allows the agent to overwrite its own policy and forge audit receipts
plugins/signed-audit-trails/skills/signed-audit-trails-recipe/references/cedar-policy.md:58
The signed-audit-trails recipe (and the matching example in plugins/protect-mcp/README.md) stores the Cedar policy (./protect.cedar), the private Ed25519 signing key (./protect-mcp.key), and the receipt log (./receipts/receipts.jsonl) inside the project root, including in its CI workflow. At the same time, the Cedar policy permits Write and Edit on all paths matching /path/to/project/* and permits Read and Bash cat* on all files without excluding governance files. A prompt-injected agent can edit ./protect.cedar to allow all future tool calls, read ./protect-mcp.key, and rewrite ./receipts/receipts.jsonl so offline signature verification still passes.
Fix: Store protect.cedar, protect-mcp.key, and receipts/ outside the project workspace via PROTECT_MCP_POLICY, PROTECT_MCP_KEY, and PROTECT_MCP_RECEIPTS, and add explicit Cedar forbid rules blocking Read, Write, Edit, and Bash access to policy files, *.key files, receipts, and .claude settings.
-
LowCedar Bash prefix permit for npm, pnpm, and yarn allows arbitrary code execution without project scripts
plugins/signed-audit-trails/skills/signed-audit-trails-recipe/references/cedar-policy.md:23
The recipe's Cedar policy permits any Bash command starting with npm*, pnpm*, or yarn* and states in its comment that these rules are only as safe as the project's scripts. However, subcommands such as npm exec -- node -e "...", pnpm dlx, or yarn node -e "..." execute inline code or remote packages without using the project's package.json scripts and without using any of the characters blocked by the policy's shell-chaining forbid rule. A prompt-injected agent can use these commands to bypass the policy's destructive-command blocks and project-directory write limits.
Fix: Replace broad npm*, pnpm*, and yarn* prefixes with exact trusted commands (such as npm test) and add forbid patterns for exec, dlx, create, init, --eval, and -e flags.
