---
title: "Matheus Theodoro"
description: "AI security engineer at Adapta. LLM red teaming, agentic security, and practitioner writing."
url: https://matheus.theodoro.dev/
---

# Matheus Theodoro

> AI security engineer at Adapta, previously founder of Avenza Security (2024 — 2026). I build and break AI systems so companies can ship them with evidence, not hope.

I red team LLMs and agentic systems, then turn confirmed failures into engineering controls. My work combines offensive security with production engineering: adversarial campaigns, judge-based scoring, and defense validation for real deployments.

## Writing

- [The Red Teamer's Guide to Agentic Defenses: 12 Systems and Where They Break](https://matheus.theodoro.dev/blog/the-red-teamers-guide-to-agentic-defenses.md): A technical survey of every major agentic defense system published in 2025-2026 — SafeAgent, AgentSentry, CoopGuard, Constitutional Classifiers, RvB, FlowGuard, SHIELD, DynaTrust, TrajGuard, AegisLLM, Stable Agentic Control, and PocketAgents — with an attacker's perspective on each.
- [FlowSteer: The Attack That Current Multi-Agent Defenses Can't See](https://matheus.theodoro.dev/blog/flowsteer-the-attack-that-mas-defenses-cant-see.md): Deep technical analysis of FlowSteer — a prompt-only planning-time attack against multi-agent LLM systems that achieves +55% malicious success rate by manipulating workflow formation itself, completely bypassing post-hoc defenses like G-Safeguard and ARGUS.
- [Constitutional Classifiers v2: What 1,736 Hours of Red Teaming Revealed About Production AI Defense](https://matheus.theodoro.dev/blog/constitutional-classifiers-v2-what-1700-hours-of-red-teaming-revealed.md): An attacker's analysis of Anthropic's next-generation Constitutional Classifiers — the two-stage cascade, linear probe screening, 0.05% flag rate, 40x cost reduction, and what the absence of a universal jailbreak actually means.

All writing: https://matheus.theodoro.dev/blog/index.md

## Experience

- **AI Security Engineer, Adapta** (2026 — Present): AI security.
- **Founder & AI Security Engineer, Avenza Security** (2024 — 2026): practical AI security practice for companies shipping LLM applications and agents.
- **Software Engineer II, Avenza Cloud** (2022 — 2026): secure cloud infrastructure and backend systems at scale.

Full history: https://matheus.theodoro.dev/experience.md

## Connect

- Email: dev.matheustheodoro@gmail.com
- GitHub: https://github.com/matheusht
- LinkedIn: https://linkedin.com/in/matheusht