Staff / Principal Offensive Security Engineer, Frontier Cyber Models

Tykhe Inc; San Francisco (CA); On-site; USD 180,000 - 280,000; Full time

Job summary

Frontier Cyber Models seeks a Staff/Principal Offensive Security Engineer to lead the development and evaluation of offensive capabilities within LSI Offense Models. You will design labs, kill chains, and training signals, working closely with ML researchers and defense engineers.

You will guide multi-stage adversary simulations, ensure ethical practice, and drive real-world applicability across enterprise-grade environments and cloud estates.

Qualifications

• 5+ years hands-on offensive security (Staff), 8+ (Principal) across at least three areas.

Responsibilities

• Own offensive capability for LSI Offense Models.
• Build Cyber World Models with realistic labs and ranges.
• Design kill chains and adversary simulations mapped to MITRE ATT&CK.
• Turn tradecraft into training signals with tools and curricula.
• Run the verify loop to triage model output and reduce false positives.
• Build and defend benchmarks for offensive evaluation.
• Feed attackers' telemetry to train Defense Models.
• Architect across security functions for estate-wide reasoning.
• Lead in field pilots, disclosures and public talks.

Skills

Hands-on offensive security

Python

Go/Rust/C/C++

LLMs and AI agents

Adversary simulation

Tools

Burp

Ghidra/IDA

AFL++, libFuzzer

CodeQL/Semgrep

C2 frameworks

Job description

Staff / Principal Offensive Security Engineer, Frontier Cyber Models

Team: Offense

Level: Staff or Principal, set by scope and track record

fix vulnerabilities before they become exploits.

detect, investigate and triage the exploits that get through. Both deliver frontier accuracy at

real-time speed, at a fraction of the cost of frontier general-purpose models on security

We are a team of security researchers, ML researchers and engineers. The people we hire turn

elite human tradecraft into the environments, rewards, evaluations and data that make our

models better every week. Learn more at lsi.inc.

The role

You are the attacker in self-play loop. You will turn what elite pentesters, red teamers and

vulnerability researchers do into environments, tasks, rewards and evaluations that make LSI

Offense Models find, prove and fix real vulnerabilities at machine scale.

This is not a consulting or internal red-team seat. Your \"targets\" are the models: you decide

which attack paths are worth teaching, design the labs they learn in, judge whether a finding is

real, and own how we measure offensive capability. You will also work directly with design

partners, running the model against their code and environments and feeding what you learn

back into training.

What you'll do

Own offensive capability for LSI Offense Models.

Set the technical direction, operating model and evaluation strategy for how the model finds, exploits, validates and

Build Cyber World Models.

Architect realistic, reproducible labs and ranges: vulnerable apps and services, cloud and Kubernetes estates, CI/CD pipelines, enterprise AD networks, and agentic / MCP-based systems. Instrument them so every step of an attack is observable and verifiable.

Design kill chains and adversary simulations.

Build multi-stage scenarios from initial access through privilege escalation, lateral movement and impact, mapped to MITRE ATT&CK, and emulate real threat-actor tradecraft rather than toy CTFs.

Turn tradecraft into a training signal.

Translate expert workflows and intuition into tools, skills, harnesses, task curricula, reward functions and expert trajectories. Define what \"proven\" means for a finding (working PoC, reachable sink, crash, verified patch).

Run the verify loop.

Triage model output, separate real vulnerabilities from noise, reverse-engineer false positives and false negatives, and close the loop so the model gets measurably better.

Build and defend benchmarks.

Extend and harden our offensive evals ranges) and build held-out sets that resist contamination.

Feed the defender.

Generate attack telemetry and campaigns that train LSI Defense Models, and work daily with the Defense engineer on purple-team self-play.

Architect across security functions.

Connect AppSec, cloud, identity, network and AI-agent security into coherent attack surfaces, so the model reasons about an estate, not a single bug.

Lead in the field and in public.

Run head-to-head pilots with design partners, coordinate responsible disclosure for the new vulnerabilities the model finds, and represent LSI's research in papers, talks and write-ups.

What you bring (must-haves)

5+ years hands-on offensive security (Staff), 8+ (Principal), across at least three of:

application security / code review, penetration testing, red teaming, adversary simulation, vulnerability research and exploit development.

Deep AppSec.

Source-level vulnerability discovery across memory-safety and logic bugs; strong on OWASP Top 10 and beyond (auth flaws, deserialization, SSRF, injection, race conditions, supply chain).

Real adversary experience.

Led red-team or adversary-emulation operations against mature enterprises or cloud estates (AWS, GCP or Azure, plus Kubernetes and identity providers), with full kill chains, not just scans.

Labs and environments.

Built ranges, vulnerable-by-design targets, CTF infrastructure or emulation platforms that others used at scale.

LLMs and AI agents in your daily work (required).

You have built or extended agents, harnesses or tool-using LLM workflows for offensive tasks and can show what they did.

You know where models fail at security reasoning and how to evaluate them.

Strong engineer.

Fluent in Python plus one of Go, Rust, C/C++ or TypeScript; you write robust tooling, not one-off scripts. Comfortable with Burp, Ghidra/IDA, fuzzers (AFL++, libFuzzer), CodeQL/Semgrep and C2 frameworks.

Judgment.

You know which bugs and attack paths matter, and you can explain that to ML researchers, CISOs and engineers alike.

Authorized, ethical practice.

Commitment to safe handling of exploits, coordinated disclosure and dual‑use risk.

RL and environments:

designed RL environments, reward functions or verifiers; worked with post-training (SFT, preference optimization, RL) on agentic tasks.

Built or led a team on AIxCC, DEF CON CTF, Pwn2Own or a top bug-bounty leaderboard; public CVEs.

Offensive testing of AI systems: prompt injection, tool and MCP abuse, agent identity and privilege, using OWASP LLM / Agentic Top 10 and MITRE ATLAS.

Built offensive automation or autonomous pentest products at a vendor or lab.

Publications or talks at Black Hat, DEF CON, USENIX Security, IEEE S&P or similar.