Trent

Research

Building the technology foundation for the AI Security Engineer.

Research-led Top researchers World-class AI research. Hands-on experience building real-world systems.
No resets Learning in live systems Agents that learn on the job, in systems where no attempt runs twice.
Compounding Security that builds on itself Each investigation keeps what it proved, so the next starts further ahead.

The research behind the judgment layer

Three areas of work shape how Trent's agents are built.

Continuous improvement
Reinforcement learning for agents that act in live systems and get better from what happened.
Customer-aligned judgment
Security judgment that aligns with customers' security policies and trade-offs.
Specialized agents
Self-managing agent systems that have to stay up, at scale.

Your AI security engineer

Agents built on this thinking, running in your environment

Weighs evidence, not severity labels
Acts within the limits you set
Builds verified context with every fix

Latest research result

How AI agents learn in systems you cannot reset.

NeurIPS 2026 arXiv 2610.03361

Follow the Winners: Conservative Policy Improvement with the Cross-Entropy Method for Critic-Free RFT

Joery A. de Vries, Neil D. Lawrence, Zhenwen Dai

Methods like GRPO train an agent by running each task several times from the same start and comparing the runs. An agent at work in a live service or a security sandbox cannot do that: each run changes the system, and an investigation may give one useful attempt.

Follow the Winners keeps the attempts it has, picks the best few from random batches of them, and learns from those in small, careful steps.

Read the paper PDF

What it found

It matches GRPO and PPO, without a critic or repeated runs.

It matches GRPO and PPO

On Search-R1's test set of over 50,000 questions, a model of 3 billion parameters reaches 35.46% average exact-match accuracy, against 33.6% for GRPO and 32.5% for PPO as their authors published. On Sokoban's puzzles it matches both.
It needs no critic and no repeated runs

No second model judges each step, and no task is run again from the same start. What it remembers sits in ordinary memory, often cheaper than a GPU's.
It holds steady when rewards are rare and extreme

In a test where a few attempts pay off hugely and most pay little, GRPO lost track of the best option, while Follow the Winners stayed steady, as DPO did.
It keeps its options open when it cannot see everything

Where two situations look the same to the agent, it stayed closer to the best strategy than GRPO and DPO, as its careful steps keep it from settling too soon.
It explains the methods around it

Seen as inference, GRPO is neutral to risk, while DPO and Follow the Winners lean toward the best outcomes, by an amount Follow the Winners keeps in check.

Why it matters

Security work happens in systems you cannot reset.

An AI security engineer works in systems that are live: a production service, a sandbox that is costly to rebuild, an investigation that gives one useful run.

Methods that need the same moment again and again do not fit there. Learning from the best of what already happened does.

That is how the research makes Trent's security better. Trent already carries forward what it verifies, a finding it proved or a correction you made. Learning from the attempts that worked is how an agent gets better at the work itself, in the systems where the work happens.

What the research is after is security that compounds: each investigation keeps what it proved, so the next starts further ahead instead of from nothing.