guest@geearr:~/services$ man geearr-ai
NAME
geearr-ai — ai & vibe-coded app security by George Rios
AI & Vibe-Coded App Security: Your AI wrote the code. Who checked it?
SYNOPSIS
geearr ai [--app] [--llm] [--agents] [--prod-ready] <app>DESCRIPTION
AI coding tools ship features fast — and they also ship hard-coded keys, missing authorization checks and databases open to the internet. Meanwhile AI agents are being handed real tools and real data. geearr audits both: the code your AI wrote, and the agents you are about to trust with production.
OPTIONS — WHAT GETS TESTED
--app · AI-built applications
- Secrets in code, client bundles and repos
- Row-level security & authorization on every route
- Test vs. live data separation
- Personal data collection, storage & retention
- Backups, outages & handoff-readiness
--llm · LLM & agent systems
- Direct & indirect prompt injection
- Tool / MCP server permission scoping
- Data exfiltration via tool calls & rendered output
- Model-output trust boundaries & sandboxing
- Tamper-evident logging of agent actions
--prod-ready · Production readiness
- Abuse, rate limiting & cost-exhaustion controls
- Hallucinated or typo-squatted dependencies
- Auth provider config: Supabase, Clerk, Auth0, Firebase
- Monitoring & alerting
- Scale & failure-mode review
OUTPUT — WHAT YOU GET
- 01Eight-area risk report: secrets, environments, access, personal data, outages, handoffs, backups, growth
- 02Fix-first list sized for a small team
- 03Agent permission model & guardrail recommendations
- 04Patches for critical issues
- 05Optional 30-day Architecture Discovery
EXAMPLES
$ geearr ai --app ./my-lovable-app service_role key found in client bundle → rotate now rls disabled on 4 tables: profiles, orders, invoices, files
$ geearr ai --agents support-bot indirect prompt injection via uploaded PDF → ticket data exfil fix: tool allowlist + human approval for outbound email
# illustrative output
CONFORMING TO
- OWASP Top 10 for LLM Applications
- OWASP Agentic AI Threats
- NIST AI RMF
- MITRE ATLAS
FAQ
[+][-]Is code generated by AI tools secure?
Not by default. AI coding assistants frequently produce working code with missing authorization checks, secrets embedded in front-end bundles, permissive database rules and outdated or non-existent dependencies. The code may function perfectly while exposing every customer record. An independent security review before launch is strongly recommended.
[+][-]What is prompt injection?
Prompt injection is an attack where untrusted text — in a web page, email, document or tool result — contains instructions that an LLM follows as if they came from the user or developer. In agents with tool access it can lead to data exfiltration or unauthorized actions. Defenses include least-privilege tools, human confirmation for sensitive actions, output filtering and isolating untrusted content.
SEE ALSO
guides: Security checklist for AI-built (vibe-coded) apps
glossary: secrets-scanning(7), least-privilege(7), row-level-security(7), vibe-coding(7), prompt-injection(7), mcp(7)
shipped: engrave(8) — The black box recorder for AI agents., identizen(8) — Accountless identity.
services: geearr-pentest(1), geearr-appsec(1), geearr-email(1), geearr-cloud(1)
EXIT STATUS
0 when every finding is closed and retested. Anything else means we keep going.
BUGS
None known. That’s what retests are for. Found one in this site? security.txt