boost
- Type
- Skill
- Scanned
- Nov 12, 2026
- Version
- b284d5c20974
- Files read
- 191
Laravel Boost is a Laravel package that installs AI coding guidelines, agent skills, browser console logging, and a local MCP server for inspecting databases, logs, routes, and Laravel documentation. Its biggest risk is an unauthenticated, CSRF-exempt browser log endpoint enabled by default in local and debug environments that lets any website the developer visits write forged entries and prompt-injection payloads into logs read by the AI agent, alongside disabled TLS certificate verification on documentation searches in local environments.
What it can reach
- Files
- Reads application logs, browser logs, package manifests, and PHP enum files, and writes AI guidelines, project rules in .ai/rules, skills, boost.json, and agent MCP configuration files.
- Network
- Registers a POST /_boost/browser-logs route on the host Laravel app, sends documentation queries and installed package names to boost.laravel.com, and downloads remote skills and audit data from GitHub and skills.laravel.cloud.
- Shell commands
- Runs local processes to detect installed AI tools, list PHPUnit/Pest tests, execute MCP tools in isolated php artisan subprocesses, and run arbitrary PHP via artisan tinker only if boost.tinker_tool_enabled is explicitly turned on.
- Secrets
- Sends a configured GitHub token (boost.github.token or services.github.token) to GitHub when downloading remote skills, and exposes database rows and log entries to the AI agent.
- Browser
- Injects JavaScript into HTML responses on web middleware routes in local or debug environments to capture console messages and uncaught page errors and post them to /_boost/browser-logs.
- Other
- Connects to the Laravel application's configured databases to read schemas, views, triggers, stored routines, and query results.
Findings
-
HighUnauthenticated, CSRF-exempt browser log endpoint allows cross-site log poisoning and prompt injection
src/BoostServiceProvider.php:147
When Boost runs (by default whenever APP_ENV=local or APP_DEBUG=true, including misconfigured staging servers), it registers POST /_boost/browser-logs without authentication, Origin checks, rate limits, size limits, or CSRF protection, and explicitly parses text/plain JSON bodies sent via sendBeacon or no-cors fetch. Any untrusted website the developer visits in their browser (or anyone able to reach the dev/debug server over the network) can silently flood storage/logs/browser.log to fill disk space or plant disguised prompt-injection instructions that are fed straight into the AI coding agent when it calls the browser-logs MCP tool. Passing an invalid log type also triggers an unhandled InvalidArgumentException that writes to the main application error log.
Fix: Verify that the Origin or Sec-Fetch-Site header matches the application origin, validate the logs array and restrict log types to an explicit allow-list, enforce payload size and rate limits, and avoid enabling the route outside local environments just because APP_DEBUG is true.
-
MediumTLS certificate verification disabled for documentation API requests in local environments
src/Concerns/MakesHttpRequests.php:20
MakesHttpRequests::client() calls withoutVerifying() whenever app()->environment(['local', 'testing']) is true, rather than only when testing or calling a local .test URL. Because developers almost always run Laravel with APP_ENV=local, every SearchDocs MCP request to https://boost.laravel.com/api/docs disables HTTPS certificate and hostname verification. An attacker on the same Wi-Fi or local network can intercept these requests to read the project's installed package list (including private package names) and search queries, and return tampered documentation containing prompt-injection instructions that the AI agent treats as authoritative Laravel docs.
Fix: Remove 'local' from the app()->environment() check so TLS verification is only skipped during automated tests or when the target URL host itself ends in .test.
-
MediumMySQL backslash-escape mismatch lets crafted SQL bypass the read-only query filter
src/Mcp/Tools/DatabaseQuery.php:120
DatabaseQuery is enabled by default and marked read-only so MCP clients can run it without user confirmation. While addPrefixToQuery() tells withoutLiteralsAndComments() that MySQL and MariaDB treat backslashes as quote escapes, isReadOnlyQuery() calls withoutLiteralsAndComments($query) with backslashEscapes left as false. If a prompt-injected agent runs a query containing '\'' on MySQL or MariaDB, MySQL closes the string literal at the second quote while the filter treats '' as a doubled quote and stays inside a string literal, hiding everything after it from keyword, semicolon, and version-comment checks. Although MySQL's SET TRANSACTION READ ONLY still blocks ordinary table modifications, hidden clauses such as INTO OUTFILE / INTO DUMPFILE (when the DB user has FILE privileges), stacked COMMIT + write statements (when PDO::ATTR_EMULATE_PREPARES is enabled), or writes on drivers without read-only transaction support (such as sqlsrv) can slip past the filter.
Fix: Require isReadOnlyQuery() to pass under both backslashEscapes=false and backslashEscapes=true (covering MySQL regardless of whether NO_BACKSLASH_ESCAPES is set), and reject connections whose driver cannot enforce read-only transactions.
-
MediumRemote skill security audit fails open, omits branch names, and auto-installs risky skills under --no-interaction
src/Console/AddSkillCommand.php:256
When installing third-party skills with boost:add-skill, the built-in security audit has three gaps: (1) if the command runs with --no-interaction (-n) — which Boost's own guidelines instruct AI agents to pass to every Artisan command — runAuditBeforeInstall() prints the audit table and returns true even when skills are flagged as Critical or High risk; (2) auditSkills() sends only owner/repo without the branch name from a /tree/<branch> URL, so an attacker's branch is audited against the repository's default branch while files are downloaded from the attacker's branch; and (3) SkillAuditor::audit() uses a 3-second timeout and silently returns an empty array on network errors or unknown skills, which is treated as safe with no warning.
Fix: Refuse to install risky skills in non-interactive mode unless --skip-audit is explicitly provided, warn and require confirmation when the audit API fails or has no results for a skill or non-default branch, and pass the target branch ref to the audit service.
-
LowNightwatch MCP setup configures unpinned npx -y mcp-remote execution
src/Install/Agents/Cursor.php:63
When boost:install configures the optional Nightwatch MCP server for Cursor, Codex, or Junie, it writes a command that executes npx -y mcp-remote with no version pin. Running npx -y on an unpinned package silently downloads and executes the latest published npm version on the developer's machine whenever the local cache misses, creating avoidable npm supply-chain exposure.
Fix: Configure native remote HTTP MCP URLs where supported by current versions of Cursor and Codex, or pin mcp-remote to a specific vetted version.
