Cyrus Nikolaidis

Cyrus Nikolaidis

Email · LinkedIn · GitHub · Scholar · Hugging Face

I build safety systems for AI, spanning evaluations, guardrails, and scalable monitoring for deployed models and products.

At Meta Superintelligence Labs, I lead work on safety evaluations and monitoring for deployed products and models. Products I've worked on safety for include Muse Spark, Meta AI, Muse Image, and the Muse personal AI agent. Before moving to MSL, I led and contributed to team efforts behind several open-source safety projects at Meta, including CyberSecEval, LlamaFirewall, Prompt Guard, and Llama Guard.

Earlier at Meta, I incubated applications of machine learning to privacy and security problems. Before that, I worked on recommendation systems, building the first version of Instagram's feed recommendations ranking and working on Facebook Search.

Selected updates

Sep 2026
Helped launch Muse, Meta's personal AI agent. We spent months stress testing and building evaluations and monitoring for agent safety and adversarial robustness to meet the demands of a major leap in agent capabilities.
Jul 2026
Helped launch Muse Spark 1.1 and the Meta Model API. Our team worked on content and agentic safety evaluations, setting up monitoring over the API, and contributing to the preparedness report.
Jul 2026
Muse Image launch — our team helped set up scalable online safety monitoring for media generation.
Apr 2026
Muse Spark launch at Meta — I worked on online monitoring strategy and implementation for the release, and supported new offline benchmarks for content safety. More in the safety and preparedness report.
Aug 2025
Spoke at Black Hat Arsenal and The Diana Initiative in Las Vegas.
Jun 2025
Joined Meta's Superintelligence Labs as an AI Research Engineer, working on safety evaluations and monitoring.
Apr 2025
Released Llama Prompt Guard 2 86M and 22M — compact models for detecting jailbreak and prompt-injection attacks.
Apr 2025
Published LlamaFirewall — a framework for deploying AI agent defenses. (Website, GitHub)
Feb 2025
Hosted a workshop and demo at AI Security Forum in Paris.
Aug 2024
Spoke at DefCon AI Village on evaluations and guardrails against prompt injection attacks.
Aug 2024
Hosted a workshop at AI Security Forum.
Jul 2024
Published CyberSecEval 3 — advancing the evaluation of cybersecurity risks in LLMs.
Jul 2024
Co-authored The Llama 3 Herd of Models paper — our team contributed cybersecurity risk measurements and system-level safety models.
Jul 2024
Launched PromptGuard and LlamaGuard 3, part of Meta's system-level safety stack for Llama.
Apr 2024
Published CyberSecEval 2 — a wide-ranging cybersecurity evaluation suite for LLMs.
Dec 2023
Published Purple Llama CyberSecEval — a secure coding benchmark for language models.