Senior Red Team Operator, AI/LLM
3 нед. назад
Страна неизвестнаSeniorRemote
aipenetration testingvulnerability assessmentcybersecurityred teaming
Senior role focused on autonomous AI-driven offensive security operations to discover and exploit vulnerabilities continuously.
О компании
- At , we're a group of engineers, researchers, and security researchers on a mission to redefine what proactive security can do in the AI era. Cyberattacks are becoming autonomous and relentless and we believe defending against threats before they materialize is one of the most powerful ways to protect the institutions the world depends on.
- We're building autonomous proactive security from the ground up, reinforcing the tradecraft of elite security practioners into purpose-built security models and agents that discover risk and remediate it before organizations are breached.
- Led by Kevin Mandia, founder of Mandiant ($5.4B exit to Google), our team brings together researchers and engineers from Google, xAI, Meta, Stanford, Georgia Tech, Waterloo, Berkeley, and MIT to reinvent security for an adversary that never sleeps.
Обязанности
- Every operator knows the ending. You found the path, chained it to access the crown jewel, and wrote it up. Then the engagement closed, the environment changed the next week, and nobody walked that kill chain again until next year's test. is built to end that cycle.
- We're building an agentic attacker swarm. These are specialized AI agents running custom models that discover, exploit, and chain vulnerabilities across external, internal, identity, cloud, and application attack surfaces. They run continuously, at machine speed, against live production environments. Until recently, nobody could put a nation-state-caliber adversary inside every network around the clock. That's the hyperattack, and it's what we do. In one recent engagement, the swarm generated more than 17 million offensive actions in three days and validated 38 attack paths. That's the scale your tradecraft will run at.
- As a Senior Red Team Operator, your tradecraft is the training signal. You'll run hands-on operations, and then you'll encode how you think into the reasoning our agents use at scale: how you pick a foothold, abuse an application's logic, pivot through identity, separate what's actually exploitable from scanner noise, and map the blast radius once you're in. You'll pressure-test full kill chains end to end, find where the swarm is naive, and teach it to move like an expert, using AI throughout your own workflow to go faster and deeper.
- The problems here are new ones:
- What does a kill chain look like when it can branch a thousand ways in parallel?
- How do you prove exploitability instead of just flagging it?
- What does a continuous hyperattack find that a two-week engagement never could?
- You'll work alongside operators who have spent decades on the front lines of red teaming and the AI engineers turning that experience into autonomous defense. It's a small, scrappy group with high ownership, and the work you ship reaches enterprises through industry leading partners. If you've ever finished an op thinking "imagine running this everywhere, all the time," this is where you get to do it.
- Run full kill chains, not checklists. Attack LLMs, agents, RAG pipelines, and the applications wrapped around them the way a real adversary would. Chain prompt injection, tool abuse, and integration flaws into paths no guardrail benchmark ever sees.
- Prove what's exploitable. Turn a jailbreak or odd model behavior into a working, reproducible exploit with clear evidence of impact. Separate real risk from refusal-bypass noise.
- Map the blast radius. Follow every foothold to its end, from a poisoned document or hijacked agent to tool calls, credentials, identity, cloud, and data. Show what an attacker could actually reach through the model.
- Hijack the agent. Abuse tool and function calling, MCP servers, plugins, memory, and multi-agent handoffs to make agents act against their owners. Target excessive agency, confused-deputy flaws, and broken trust boundaries.
- Poison the pipeline. Attack the full AI supply chain, including training and fine-tuning data, embeddings and vector stores, model weights and serialization formats, and the MLOps infrastructure that ships them.
- Read code like an attacker. Review agent frameworks, orchestration logic, and model-serving code for injection paths, unsafe output handling, and insecure design, using AI-assisted analysis to cover more ground faster.
- Take models apart. Use black-box probing, adversarial input generation, model extraction, and membership inference to understand how a model really behaves and what it leaks.
- Train the swarm. Translate your AI attack tradecraft into the techniques, decision logic, and evaluation data that make our agents operate like expert operators at machine speed.
- Break the machine. Pressure-test the swarm's own models and agents, spot where they're naive, manipulable, or wrong, and make the human-in-the-loop calls that keep autonomous operations safe and accurate.
- Build what's missing. Write scripts, harnesses, and tools that automate AI attacks at scale, and treat AI as a daily force multiplier in your own workflow.
- Tell the attack story. Deliver reports and debriefs that walk engineers, executives, and counsel through the attack narrative, the business risk, and the fix.
- Stay ahead of the threat. Research emerging AI attack classes, from indirect prompt injection to agent-to-agent exploitation, and feed what you learn straight into the platform.
- Own your work. Balance concurrent operations and internal research with startup-level autonomy, shifting priorities as fast as the threat does.
Требования
- Deep, hands-on AI offense. LLM and agent exploitation beyond the OWASP Top 10 for LLM Applications, including direct and indirect prompt injection, jailbreak chains, system prompt extraction, and data exfiltration through model outputs.
- Depth in at least one more AI attack surface: Agentic systems: tool and function-call abuse, MCP and plugin exploitation, memory poisoning, and multi-agent trust boundaries.
- RAG and data pipelines: vector store and embedding attacks, retrieval poisoning, and cross-tenant data leakage.
- Adversarial ML: evasion, model extraction, membership inference, and training or fine-tuning data poisoning.
- AI supply chain and infrastructure: malicious model artifacts, unsafe serialization, model registries, and inference and MLOps platforms.
- Application security fundamentals. Most AI attacks land through the app around the model. You bring strong web and API exploitation skills, including authentication and authorization bypass and injection chains.
- An attacker's toolkit and the instinct to go beyond it. Fluency with tools like Burp Suite, Caido, Garak, PyRIT, promptfoo, and the Adversarial Robustness Toolbox, plus the judgment to know when the tool is wrong.
- Systems understanding that makes chaining possible. Strong grasp of how LLMs, agent frameworks, and model-serving stacks work under the hood, and how one weakness leads to the next.
- Comfort everywhere you land. Proficiency across Linux, Windows, and macOS and major cloud AI platforms, with scripting in Python, Bash, or similar.
- An AI-forward way of working. You already use AI tools to speed up recon, analysis, or tooling, and you're excited to push them much further.
- Sharp human-in-the-loop judgment. You absorb new information quickly, make time-sensitive calls under pressure, and can explain your reasoning to safety and engineering teams.
- A builder's, startup mindset. You're scrappy, self-directed, and comfortable with ambiguity, and you'd rather ship something useful than wait for a process.
- Clear communication with clients. You can translate a complex AI attack chain for engineers, executives, and legal counsel during engagements and remediation debriefs.
- Logistics. Must be eligible to work in the United States without sponsorship. Ability to travel up to 20% for on-site engagements or client meetings.
Условия
- Remote in United States
- 🏥 Full Health, Dental, & Vision Coverage
- 📈 Meaningful Equity Ownership
- 🥙 In-Office Meals
- ✂️ Haircuts at the Office 🎉 Company Sponsored Conferences & Events
- 💸 401(k), HSA, and FSA Plans
- 🌴 Flexible PTO
Другое
- You've built AI systems, not just broken them. Hands-on work with LLM fine-tuning, agentic frameworks like LangChain or AutoGen, or ML pipelines gives you an insider's view of where they fail.
- You've shipped offensive AI tooling. Prompt injection harnesses, jailbreak fuzzers, Burp or Caido extensions for LLM traffic, or open-source AI red team projects others actually use.
- You hunt in public. AI bug bounty findings, CVEs in AI frameworks or model-serving platforms, AI CTF wins, or conference talks at venues like DEF CON's AI Village.
- You keep sharpening. Active time on AI security challenges like Gandalf, Crucible, HackAPrompt, or PortSwigger's LLM attack labs.
- You know the frameworks. Working familiarity with MITRE ATLAS, the OWASP Top 10 for LLM Applications, and the NIST AI Risk Management Framework.
- Relevant certifications, such as OSWE, OSAI, Burp Suite Certified Practitioner, GWAPT, or emerging AI security credentials.