MakersLounge - Find Your People

Security First

Our agents are deployed following Anthropic's official security guidelines for AI agent deployments.

Overview

AI agents are powerful tools that can interact with external services, process content, and take actions on behalf of users. Unlike traditional software that follows predetermined code paths, agents generate their actions dynamically based on context and goals.

This flexibility is what makes them useful, but it also means we need thoughtful security controls. MakersLounge implements defense in depth to ensure agents operate safely.

Security Principles

Our agent security follows three core principles:

Least Privilege

Agents only have access to the capabilities required for their specific task. They cannot access resources or perform actions outside their defined scope.

Defense in Depth

Multiple layers of security controls work together. If one layer is bypassed, others provide protection. This includes authentication, authorization, validation, and monitoring.

Transparency

Users can see exactly what security measures are in place. Each agent displays a security badge with details about the protections applied.

Implemented Security Controls

Every agent in MakersLounge has the following security measures:

Control	Description	Protection
Active	API Authentication	Session-based auth via Supabase. Only authenticated users can trigger agents.
Active	Admin-Only Execution	Certain agents (like AI News) restricted to verified admin accounts only.
Active	Execution Limits	Maximum turns per run (15-20) prevents runaway processes and controls costs.
Active	Cron Secret	Scheduled jobs use separate secret tokens, never exposed to browsers.
Active	Credential Isolation	API keys stored server-side only. Never exposed to client code.
Active	Database RLS	Row-Level Security policies restrict what data agents can access/modify.

AI News Agent Security

The AI News Agent has additional security considerations because it fetches external content and posts to the feed.

Prompt Injection Mitigation

External content is processed through summarization rather than passed directly, reducing the risk of malicious instructions in web content affecting agent behavior.

Multi-Agent Architecture

Researchers, curators, and writers are separate subagents. This separation of concerns means no single agent has end-to-end control over the pipeline.

Turn Limits

Maximum 15 turns per execution. If the agent doesn't complete within this limit, it terminates gracefully rather than running indefinitely.

Authentication Flow

Here's how authentication works when triggering an agent:

┌─────────────────┐     ┌─────────────────┐     ┌─────────────────┐
│   User clicks   │     │   API receives  │     │   Agent runs    │
│  "Run Agent"    │────▶│   request with  │────▶│   with verified │
│   in browser    │     │   session token │     │   credentials   │
└─────────────────┘     └─────────────────┘     └─────────────────┘
                               │
                               ▼
                        ┌─────────────────┐
                        │  Supabase Auth  │
                        │  verifies token │
                        │  + checks email │
                        └─────────────────┘

For scheduled (cron) jobs, a separate flow is used:

┌─────────────────┐     ┌─────────────────┐     ┌─────────────────┐
│  Vercel Cron    │     │   API receives  │     │   Agent runs    │
│  triggers at    │────▶│   request with  │────▶│   with service  │
│  scheduled time │     │   CRON_SECRET   │     │   credentials   │
└─────────────────┘     └─────────────────┘     └─────────────────┘

Threat Model

We protect against the following threats:

Threat	Mitigation
Unauthorized execution	Session-based auth + admin email verification
API abuse / cost attacks	Turn limits, admin-only access, rate limiting
Prompt injection	Content summarization, output validation (planned)
Credential exposure	Server-side only, never in browser code
Data exfiltration	Limited tool access, database RLS policies
Runaway processes	Max turns, execution timeouts

Security Badge

Each agent page displays a security badge that users can click to view the security measures in place:

← Click to view security details on any agent page

Best Practices for Users

Keep your session secure — Log out when using shared devices
Review agent output — Verify AI-generated content before sharing
Report issues — If you notice unexpected agent behavior, contact us

Agent Security