Cybersecurity Specialist — AI Safety Evaluation
Summary
Author Dockerized AI red-teaming evaluation tasks (agent instructions, deterministic verifiers, oracle solutions) that probe how LLM agents misbehave — prompt injection, jailbreaks, reward hacking. 100% remote, paid per accepted task, starts with a take-home assignment.
Compensation: No equity
About Solumn AI
Solumn AI builds the safety and assurance layer for frontier AI labs. We create the sandboxed evaluation environments that leading AI teams use to measure — and improve — how their models behave when it matters most. If capability asks "what can the model do?", we ask "what happens when it fails, and how do we prevent it?"
Why this matters: AI is crossing the threshold where security and assurance stop being optional. We're building the firm that does this work to a benchmark standard, and we're growing the specialist pool behind it.
The role
We're looking for cybersecurity specialists to author sandboxed evaluation environments (RLEs) that measure how AI agents behave on real security scenarios — and whether they can be pushed to cross a line they shouldn't. All work runs in sealed, isolated sandboxes for safety research. This is AI red teaming and adversarial ML, not traditional application security or VAPT.
This role is for: AI Security Engineer · LLM Security Engineer · AI Red Teamer · AI Safety Engineer · ML Security Engineer · Adversarial ML Researcher.
What you'll build
Each task you author is a complete, Dockerized package:
Agent-facing instruction that defines scope and deliverables without revealing the solution Containerized environment with the target system, dependencies, and tools, fully reproducible Deterministic verifier — a hidden grader that passes only on a genuine, replayable result and accepts alternative valid solutions Oracle solution — your own working solution that validates the task and demonstrates the intended behavior on the target model
Focus areas
Adversarial probing of agentic coding and security models · prompt injection and tool-call abuse · jailbreaks and model misuse · unsafe autonomous actions · data handling and scope-escape by agents · reward-hacking and eval-gaming detection · adversarial robustness and red-team methodology.
What we're looking for Hands-on experience probing or evaluating LLMs and agentic systems — red teaming, jailbreak/prompt-injection research, adversarial ML, or AI safety evaluation A security mindset applied to AI systems specifically — you understand how agents fail, get manipulated, or game their objective The judgment to catch a result that passes illegitimately — a non-replayable result, a gamed verifier, or an agent that reached the answer the wrong way Comfort with Docker, Git, and writing a deterministic verifier Ability to read a multi-step agent trajectory and judge whether it genuinely succeeded Nice to have
Published red-team reports, adversarial-ML research, or model-evaluation write-ups · experience with LLM-as-judge or agentic-eval frameworks · exposure to reward modeling or RL environments.
How it works
Selection includes a short take-home assignment with one feedback and revision round, then an ongoing per-task engagement. 100% remote and global, paid per accepted task, immediate start.
To apply
Share one piece of relevant work — a red-team report, model evaluation, jailbreak/prompt-injection or adversarial-ML write-up — and briefly, what the weakness was and how you proved it. Include a link if you can.
