GenAI Security Researcher

As a GenAI Security Researcher, you will attack the most capable generative AI models in the world before anyone else does. You will research frontier LLMs and the agents built on them, find where their safety and security controls break, and turn those findings into attacks, datasets and evaluations that make the next generation of models harder to exploit.

This role sits at the meeting point of two disciplines. You bring the mindset of a cybersecurity researcher (reverse engineering a system you don't fully see, chaining small weaknesses into real impact, hacker mindset) and apply it to models whose behavior is probabilistic, multimodal and increasingly autonomous. You will work hands-on, publish internally and externally, and help shape how Alice tests the frontier.

Responsibilities

• Conduct hands-on adversarial research on frontier GenAI models (text, code, multimodal and reasoning models) to uncover novel jailbreaks, prompt-injection vectors, guardrail bypasses and policy-evasion techniques.
• Research agentic attack surfaces: tool use, MCP and plugin ecosystems, browsing and computer-use agents, coding agents, and multi-agent pipelines, including indirect prompt injection, data exfiltration and tool misuse.
• Assess model capabilities in offensive cyber domains (vulnerability discovery, exploit development, malware, lateral movement) and help evaluate how much real-world uplift a model provides to an attacker.
• Translate classic offensive security techniques (privilege escalation, credential abuse, SSRF, injection classes, supply-chain attacks) into attack scenarios against LLM-based systems.
• Design and build adversarial datasets, attack taxonomies and evaluation scenarios that are realistic, reproducible and correctly graded.
• Track the evolving threat landscape: new model releases, published attacks, academic research and in-the-wild abuse, and turn them into actionable research directions.
• Automate attack generation and testing using Python, LLM-assisted tooling and agent frameworks.
• Communicate findings clearly to customer deliveries, research partners and internal teams through reports, write-ups, talks and responsible disclosure.-
• Work closely with red-teamers, data scientists and engineers to feed research into Alice's eco-system practices and knowledge share.

Must Have

• 5+ years of hands-on cybersecurity research experience in one or more of: offensive security, red teaming, vulnerability research, reverse engineering, detection research or exploit development.
• 1+ years of hands-on work (formal or a-formal) attacking or evaluating GenAI systems: jailbreaks, prompt injection / indirect prompt injection, guardrail bypass, or LLM red teaming.
• Solid understanding of how modern LLMs work and are deployed: training and alignment basics (RLHF, safety tuning), system prompts, context windows, tool calling, RAG and agent architectures.
• Strong attacker mindset: the ability to look at an unfamiliar system, model its trust boundaries and find creative ways to break them.
• Strong coding skills and comfort building your own tooling and automation.
• Ability to work independently on open-ended research problems, from hypothesis to proof of concept to clear write-up.
• Excellent written and verbal communication in English.

Preferred Skills That Set You Apart

• Published AI security research, CVEs, conference talks (e.g., DEF CON AI Village, Black Hat, BSides) or public jailbreak / prompt-injection write-ups.
• Experience red teaming production AI agents or AI-powered applications, or participation in AI bug bounty programs.
• Background in threat intelligence, incident response or detection engineering.
• Familiarity with AI security frameworks such as OWASP Top 10 for LLMs, MITRE ATLAS and NIST AI RMF.
• Experience building adversarial benchmarks or evaluation harnesses for LLMs.
• Working knowledge of ML fundamentals or experience with open-weight models (fine-tuning, inference, interpretability).
• Offensive security certifications (e.g., OSCP, OSWE, OSED, OSCE3, GREM) or strong CTF track record (HackTheBox, Pwn2Own and similar).