---
title: "Launching Elliot-1 – a model for offensive cyber tasks"
description: "The first of a family of machine learning models that make cyber intelligence accessible to everyone."
date: "2026-11-17"
authors: "Nick Vernij"
url: "https://www.tryelliot.app/blog/launching-elliot/"
---

# Launching Elliot-1 – a model for offensive cyber tasks

The most capable models on the cyber frontier have an invite-only structure and high costs. This limits who can use these models to protect themselves, and how often they can run.

Elliot-1 is an adaptation of DeepSeek V4.1 Flash, designed to run in our cyber harness at great efficiency.

Through weight-programming and abliteration we removed layers responsible for steering the model away from completing offensive cyber tasks.

We’re experimenting with incorporating internal reinforcement learning environments into the training process, adapted from CVEs and real-world scenarios. The reward structure of these environments incentivizes the model to use fewer tokens while solving the task.

## We’re opening up Elliot for everyone today

White-box pentesting with Elliot is open for everyone to use today.

Connect your GitHub repo, and get a scan for free. Elliot will map your codebase, find potential security issues, and build a working PoC before pushing them to your team.

## Why we’re training custom models

Tasks with verifiable outputs tend to be a function of compute spend. The more compute, the more results. This is true for autoresearch in machine-learning engineering, maths and cyber.

That’s why we are making access to cyber intelligence as abundant as electricity by pushing the cost-per-task to zero.

If you’re building a business on the internet today, attackers will target you. Staying safe should never be about who spends the most money.

### Benchmarks

In our internal benchmark in two categories, a capture the flag style sandbox and a set of real-world CVEs from popular open-source projects, Elliot measured 2 percentage points higher than its base model while costing 13.2% less.

During evaluation, Elliot has uncovered 24 vulnerabilities in open-source repositories, including high severity ones. We disclosed these to the maintainers and hope to share more about these soon as the parties involved roll out their fixes.

When we divide the costs of these runs by the number of vulnerabilities found, this comes down to less than $10 per vulnerability reported.

### Training

Most gains come from the model having less doubt whether to perform a task, iterating on ethics in its reasoning traces, as well as weight programming in the areas that are responsible for reiterating and confirming ideas and implementations.

This is a preview model. We spent about $1,500 in H200 compute over the span of 3 weeks evaluating DeepSeek’s weights looking for leads that push the cyber frontier. With the early results we’ve seen in weight programming and reinforcement learning environments, we’re optimistic we can push the frontier even further.

By keeping training efficient, we will be able to build on top of better frontier models quickly after their weights are released.

### Safety

This is not a model everyone should use. During evaluation, in the reasoning traces we found evidence of the model contemplating breaking out of its sandbox, as well as trying to hack the verifiers to get a good result.

We believe this is a side-effect of steering the model’s weights around ethics and guardrails as well as running the model in a swarm configuration. Swarms tend to ‘convince’ each other reward-hacking and jailbreaking is a good idea.

### Release

That’s why we’re releasing this model in our own harness first. Everyone can onboard and affordably scan their codebase with Elliot’s cyber capabilities.

We hope to offer inference and onboard select teams working on security problems at the start of next year as we evaluate the model’s safety and keep improving its capabilities.

---

From the [Elliot blog](https://www.tryelliot.app/blog/). Elliot finds exploitable vulnerabilities in GitHub repositories.
Run a free scan: https://dashboard.tryelliot.app/sign-up. Agents can sign their person up with https://www.tryelliot.app/auth.md.
