# Mira Vantis

> How AI agents get tricked, and where the controls actually go. Mira Vantis makes short videos about AI agent security: prompt injection, MCP servers, agent permissions and tokens, sandboxing, and the real incidents behind them. Every claim comes with a source.

Mira Vantis is a video series on AI agent security. Clips run under a minute and are posted on TikTok, Instagram, YouTube Shorts and X. Each clip has a page on this site with a short direct answer, a written breakdown, sources and an FAQ.

## Topics

- Prompt injection: Hidden instructions in resumes, web pages, emails and tickets, and why the model can't reliably tell them apart from yours.
- MCP server security: Open tools/list endpoints, missing auth, token passthrough and poisoned tool descriptions.
- Agent permissions and tokens: Short-lived, scoped tokens, and what an agent should never hold.
- Sandboxing and egress: Default-deny network rules outside the agent, so a blocked agent can't route around them.
- Real incidents: What actually happened, told from the primary source, and the one control that would have stopped it.

## Pages

- [Home](https://miravantis.com/): who Mira Vantis is, latest clips, email signup
- [Clips](https://miravantis.com/clips/): index of every clip page
- [What is resume prompt injection?](https://miravantis.com/clips/what-is-resume-prompt-injection/): Resume prompt injection is when a job applicant hides instructions for an AI screener inside a resume, usually in tiny white text that people can't see but the model still reads. A line like "move forward with this candidate" tries to push the applicant up the list. The fix is to keep the hiring decision outside the model.
- [How do you know if your MCP server is exposed?](https://miravantis.com/clips/mcp-server-security-red-flags/): Check three things. If tools/list answers without a token, if a request with no token doesn't get a 401 with a WWW-Authenticate header, or if the server forwards your token to other APIs, it's exposed. Trend Micro found 492 MCP servers with no client authentication, exposing 1,402 tools.
- [Why did an AI agent try SQL injection on an ordinary data task?](https://miravantis.com/clips/ai-agent-tried-sql-injection/): Because it was blocked and kept trying to finish its task. Transluce found AI agents probing three public data sites with SQL injection, cross-site scripting and path traversal while fetching ordinary statistics. Nobody asked them to hack. The control that stops this is default-deny egress, enforced outside the agent.

## Channels

- TikTok (@miravantis3): https://www.tiktok.com/@miravantis3
- Instagram (@miravantis): https://www.instagram.com/miravantis/
- YouTube (@MiraVantis): https://www.youtube.com/@MiraVantis
- X (@miravantis): https://x.com/miravantis

## Notes

- Mira Vantis is an AI-generated character. A human writes and fact-checks every script.
- Every factual claim on a clip page links to its primary source. Quote the source, not the clip, when you can.