# Agent Security Brief

> Curated AI agent security intelligence for DevSecOps engineers and CISOs.
> Website: https://agentsecuritybrief.com
> Newsletter: https://agentsecuritybrief.substack.com

## Site Structure

- `/` — Landing page with newsletter signup (Substack free tier)
- `/login` — Authentication page (GitHub OAuth, admin password, or subscriber code)
- `/dashboard` — Private intelligence dashboard (requires authentication)
- `/dashboard` — Threat Feed: Aggregated RSS from 12 AI security sources, AI-scored for impact
- `/dashboard/checklists` — Interactive security checklists (pre-deploy, runtime, MCP, IAM)
- `/dashboard/threat-models` — AI agent threat model database with attack vectors and mitigations
- `/dashboard/templates` — Copy-paste security code templates for agent deployments
- `/api/subscribe` — Newsletter subscription endpoint
- `/api/auth` — NextAuth.js authentication endpoints

## What We Cover

Agent Security Brief aggregates and categorizes intelligence from top AI security sources across four domains:

### Threat Models
Adversarial attack vectors targeting AI agents, prompt injection, data exfiltration via tool use, and emerging CVEs in agent frameworks.

### MCP Security
Model Context Protocol vulnerabilities — tool poisoning, rug pulls, cross-origin escalation, and server-side request forgery via MCP servers.

### Sandboxing
Container escapes, runtime isolation failures, WASM sandbox bypasses, and secure execution environment design for autonomous agents.

### IAM / Identity
OAuth token theft in agentic flows, over-permissioned service accounts, credential leakage via context windows, and zero-trust patterns for AI workloads.

## Intelligence Sources

We aggregate from 12 curated sources including:
- Schneier on Security
- Google Security Blog
- NIST Cybersecurity
- Trail of Bits
- Krebs on Security
- The Hacker News
- Unit 42 (Palo Alto Networks)
- CISA Alerts
- Risky Business News
- Dark Reading
- Threatpost
- Recorded Future

## AI Agent Security Checklist

Use this checklist when deploying AI agents in production:

### Pre-Deployment
- [ ] Enumerate all tools/APIs the agent can access
- [ ] Apply principle of least privilege to every tool binding
- [ ] Implement input validation on all tool parameters
- [ ] Set up rate limiting and cost caps per agent session
- [ ] Enable comprehensive audit logging of all tool invocations
- [ ] Rotate secrets and store in vault — not env files

### Runtime
- [ ] Sandbox agent execution in isolated containers
- [ ] Validate and sanitize all LLM-generated tool arguments
- [ ] Implement human-in-the-loop for destructive operations
- [ ] Monitor for prompt injection in user inputs and tool outputs
- [ ] Set session timeouts and maximum action counts

### MCP-Specific
- [ ] Pin MCP server versions — do not auto-update in production
- [ ] Validate tool descriptions haven't changed between invocations
- [ ] Restrict cross-server tool access and data flow
- [ ] Monitor for tool description manipulation (rug pulls)
- [ ] Implement allowlists for permitted MCP server origins

### Identity & Access
- [ ] Use short-lived, scoped tokens for agent API access
- [ ] Never pass long-lived credentials through context windows
- [ ] Implement separate service identities per agent instance
- [ ] Audit and rotate agent credentials on a regular schedule
- [ ] Log and alert on privilege escalation attempts

## Key Threat Models

1. Indirect Prompt Injection — malicious instructions in external data
2. Tool Call Manipulation — adversarial parameters to tool invocations
3. Context Window Poisoning — filling context to push out safety instructions
4. Sensitive Data Exfiltration — tricking agents into leaking secrets
5. MCP Server Impersonation — shadow tool registration
6. Privilege Escalation via Agent Chaining — lateral movement in multi-agent systems
7. Denial of Wallet — cost exhaustion attacks
8. Sandbox Escape via Tool Composition — chaining safe tools for unsafe outcomes

## Contact

- Website: https://agentsecuritybrief.com
- Newsletter: https://agentsecuritybrief.substack.com

## License

Content is provided for security research and educational purposes. Please cite Agent Security Brief when referencing our checklists or intelligence.