Senior AI Researcher

Assail | AI Engineering | Reports to VP of AI Engineering | Remote-friendly, Boston HQ

About Assail

Assail builds autonomous offensive security. Our platform, Ares, finds vulnerabilities in production systems by reasoning about them the way an experienced attacker would — chaining flaws across APIs, web applications, and mobile surfaces to surface the exploits that scanners miss and human testers run out of time to find.

We train our own models. Dagger is our 14B-parameter offensive security model, fine-tuned for vulnerability discovery and exploit reasoning. Javelin is our co-evolutionary self-training architecture, where attacker and defender models train against each other to push capability further than either could reach alone. The research surface is wide open, the domain is consequential, and the work ships into a platform that's actively used against hardened enterprise targets.

The Role

We're hiring our first dedicated AI Researcher to advance the core models powering Ares. You'll work alongside our VP of AI Engineering and a small AI engineering team, with direct collaboration with our CEO — a researcher and practitioner with 26 years of offensive security experience, contributions to the OWASP API Security Top 10, and a permanent exhibit at The Mob Museum.

This is a research role, not an applied ML role. You'll own original research on offensive security agents — how they reason, plan, use tools, and operate autonomously over long horizons. You'll design experiments end-to-end, build the evaluation infrastructure the field doesn't yet have, and translate research wins into capability that ships.

The feedback loop is fast and adversarial. Research that proves out goes into production. Research that doesn't gets killed quickly so the next bet can start.

What You'll Do

• Drive original research on offensive security agents — reasoning, planning, tool use, and autonomous long-horizon operation

• Advance Dagger's post-training pipeline: supervised fine-tuning, RL from verifier signals, LoRA adaptation, and evaluation against adversarial benchmarks

• Extend Javelin's co-evolutionary self-training architecture: curriculum design, self-play dynamics, and reward modeling for security-specific outcomes

• Design and execute experiments end-to-end, from hypothesis through writeup

• Build internal evaluation harnesses that measure capability rigorously, where no public benchmark exists

• Translate research into production handoffs to AI Engineering — model cards, deployment notes, and known failure modes

• Contribute to Assail's external research voice through papers, talks, responsible disclosures, and technical writing

• Collaborate with engineering teammates on research methodology and experimental design

What We're Looking For

You don't need every item on this list. We care more about depth where you have it than breadth where you don't.

Core experience that matters most:

• Original ML research output — published papers, widely cited preprints, significant open-source releases, or shipped research that materially advanced a production system

• Hands-on post-training experience with language models at the 7B+ parameter scale, end-to-end ownership of a pipeline including data, training, and evaluation

• Direct work with at least one of: RL from verifier or reward signals, preference optimization (DPO/IPO/KTO), or supervised fine-tuning with synthetic data pipelines

• Experience with agentic LLM systems — tool use, multi-step reasoning, planning, or long-horizon execution

• Ability to design evaluation that measures real capability and avoids contamination or specification gaming

• Strong Python and PyTorch, with experience in distributed training at multi-GPU scale

• Clear technical writing — research memos, experiment writeups, papers, or equivalent

Helpful but learnable here:

• Working knowledge of offensive security fundamentals (we'll teach you the rest if you bring strong ML depth)

• Prior work on code-generating or code-reasoning models

• Experience with sparse, delayed, or expensive reward signals in RL

• Research on robustness, adversarial ML, or red-teaming of language models

• Familiarity with long-horizon agent benchmarks (SWE-bench, Cybench, WebArena, or similar)

Things we deliberately don't require:

• A PhD. Track record matters more than the credential. If your work demonstrates the capability, the degree is secondary.

• A security background. Strong ML researchers can develop security depth here, and we'll support you in doing it.

• A specific number of years. Senior is a function of judgment and output, not a count.

What This Role Will Teach You

• How to train and post-train capable models in a narrow, high-stakes domain

• How to design evaluation that holds up to scrutiny when no benchmark exists yet

• How agentic systems behave under adversarial conditions — including failure modes that don't appear in benign settings

• The full offensive security stack — API, web, and mobile — at a depth most ML researchers never reach

• How to make publication and disclosure decisions for dual-use research

• How research moves from hypothesis to production in a small team where the handoff is measured in days

What We Offer

• Competitive base salary and meaningful early-stage equity

• Comprehensive health and dental coverage

• Unlimited paid time off, including parental leave

• Conference, publication, and continued learning budget — we want you engaged with the research community

• The chance to work on a problem that matters, with people who care about doing it well

Senior AI Engineer, Ares Platform

Team: Ares AI Engineering Reports to: Ilir Osmanaj, VP of AI Engineering Location: Boston, MA (hybrid) or remote with overlap to ET working hours

Position summary

The Senior AI Engineer is a core builder on the team responsible for the agents and models that power Ares — Assail's autonomous offensive security platform for APIs, web applications, and mobile applications. This role works directly on Ares' named-agent architecture (Polemos, Hermes, Enyo, Momos, Dolos, Themis, Aletheia, Argus, Kratos), the model powering Ares, and the Javelin co-evolutionary self-training loop. The engineer will ship capabilities that move the platform forward across exploit chaining, multimodal vision, mobile coverage, self-improvement, and customer-facing accuracy.

Core tasks

• Agent development. Design, implement, and continuously improve the behavior and prompting of Ares' named agents, including orchestration patterns, hand-offs, planning loops, tool use, and shared memory.

• Model training and fine-tuning. Contribute to the model powering Ares across data curation, SFT, preference optimization (DPO/GRPO-style), and evaluation. Own pieces of the training pipeline from dataset construction through eval.

• Javelin loop. Extend the co-evolutionary self-training system that lets Ares learn from its own engagements and improve over time.

• Self-improvement systems (ARES-420 and successors). Build false-positive detection, tiered skill learning (suppression rules, agent directives, code-patch proposals), and the infrastructure that routes proposed changes through human approval and back into the platform.

• Evals. Design rigorous, security-specific evaluations covering OWASP Top 10 coverage, exploit chaining, finding accuracy, and agent reliability. Track performance over every model and agent change.

• Multimodal and platform expansion. Contribute to vision capabilities, mobile (iOS/Android) coverage, and BYOK support shipping in Sidewinder and beyond.

• Production reliability. Own latency, cost, observability, and failure-mode analysis for agents running in customer engagements. Partner with the platform team on Kubernetes-based deployment.

• Customer-facing accuracy. Contribute to the live accuracy gauge and other surfaces where model and agent quality is exposed to customers.

Must-have skills

• 5+ years building production ML/AI systems, with at least 2 years working directly on LLMs or LLM-powered agents.

• Deep Python; strong, production-grade engineering practices (testing, code review, observability).

• Hands-on fine-tuning experience: SFT, preference optimization (DPO, GRPO, RLHF/RLAIF), data curation, and synthetic data generation.

• Strong grasp of transformer architectures and the modern training stack (PyTorch, Hugging Face, DeepSpeed or FSDP, accelerate).

• Experience designing and shipping multi-agent or tool-using LLM systems in production — not just demos.

• Rigorous eval design: building harnesses, tracking experiments, and making model/agent decisions based on data rather than vibes.

• Inference optimization experience: vLLM or TensorRT-LLM, quantization, throughput/latency tradeoffs.

• Comfort with retrieval pipelines, vector stores, and structured memory for agents.

• Kubernetes and containerized deployment fluency.

• Genuine interest in offensive security and the ability to ramp quickly on OWASP Top 10, API security, web app pentesting, and mobile pentesting concepts. Direct offensive security background is a strong plus but not required.

Nice to have

• Offensive security background: OSCP/OSWE/OSWA, CTF, bug bounty, or prior red team work.

• Research publications at NeurIPS, ICML, ICLR, USENIX Security, IEEE S&P, Black Hat, or DEFCON.

• Open source contributions to agent frameworks or LLM tooling.

• Experience with adversarial ML or red-teaming AI systems.

• Familiarity with mobile app reverse engineering or binary analysis.