Tortilya
In five days, one researcher and Claude Opus 4.7 broke the most capable LLMs on the planet — twenty-four different ways — then turned every failure into a lesson, a defense, and a public certification course.
Core trick
Tests agents by planting attacker-controlled instructions in the data they read, then records each bypass as a reproducible demo with a defensive playbook.
Description
In five days, one researcher and Claude Opus 4.7 broke the most capable LLMs on the planet — twenty-four different ways — then turned every failure into a lesson, a defense, and a public certification course.
This project started as adversarial research, then evolved into defensive playbooks, and finished as an enterprise learning & certification platform — all targeting one question: How can organisations stay safe in this new world full of AI agents?
inevolin/agentic-ai-safety-and-security-program★ 141 forksTypeScriptCreated Apr 22, 2026 · last push May 21, 2026
README excerpt
<div align="center"> # Agentic AI Safety & Security Program ### Adversarial research, defensive playbooks, and a full certification platform — all targeting one question: > **How do production AI agents *actually* fail when an attacker controls the data they read?** []() []() []() []() []() [](#contributing--we-want-collaborators) **[🌐 Live Training Platform → ai.nevolin.be](https://ai.nevolin.be/)** · **[Findings](attacks/demos/FINDINGS.md)** · **[Mitigations](docs/mitigations/ai-agent-security-mitigations.md)** · **[Methodology](METHODOLOGY.md)** · **[Attack Demos](attacks/demos/)** [](https://ai.nevolin.be/) </div> > ⚠️ **Research artifact.** Every attacker domain, vendor name, and portal URL inside `attacks/`, `docs/mitigations/`, and `sources/` is a placeholder used for defensive