The agent keeps going.

Claude Code, Codex and OpenCode on abliterated models. They finish the fuzzers, scanners and exploit checks in your backlog instead of stopping to decline.

Request access Read the guide

Where aligned agents stop.

Security-adjacent tasks sit in every backlog. An aligned model treats them as threats and ends the run; an abliterated one treats them as work.

Write a fuzzer for our upload parser and run itDeclinedDonePort-scan the staging subnet and diff it against the inventoryDeclinedDoneReverse the license check in our own desktop buildDeclinedDoneGenerate payload variants to test our WAF rulesDeclinedDoneDraft the phishing simulation for security awareness weekDeclinedDoneReproduce the deserialization bug from the pentest reportDeclinedDone

One minute, any harness.

Point your agent at XField, set the model, and keep everything else: commands, hooks, MCP servers, habits.

Claude Code Codex OpenCode

export ANTHROPIC_BASE_URL="https://xfield.cloud.brutal.ai/api"
export ANTHROPIC_AUTH_TOKEN="$XFIELD_API_KEY"
export ANTHROPIC_MODEL="deepseek-v4.1-flash-1017-uncensored"
export ANTHROPIC_DEFAULT_HAIKU_MODEL="glm-5.3-flash-uncensored"

claude

What an agent needs from a model

Picking a model for agents Tool calls

Reliable function calling on Messages, Responses and Chat Completions, so the agent loop never breaks.

1M tokens

Room for the whole repository map, the diff and the logs in a single session.

Cached input

Agent turns resend the same prefix. Cached tokens cost a fraction, so long sessions stay cheap.

Effort control

Low effort for edits, max for the hard bug. Set it per request or per session.

Questions

Which model should my agent use?

DeepSeek V4.1 Flash as the main model: it leads our models on coding benchmarks and stays cheap at agent volumes. GLM 5.3 Flash as the lighter model for titles and summaries. Switch to GLM 5.3 when you need very long outputs or a second opinion.

Do I lose anything compared to my usual model?

Your tools, prompts, slash commands and hooks keep working: the agent talks the same API. The difference is the model behind it, which answers instead of declining.

Is my code stored?

No. Prompts and completions pass through without being stored. Data handling.

Other use casesSecurity researchMalware analysis, reverse engineering and detection engineering, with a model that reads the sample.Red teamingAttack paths, read-only checks and reports for authorized engagements, plus test cases for your own AI agents.Trust and safetyWrite and label examples of scams, abuse and harassment so your classifiers learn what to catch.

Keep your stack. Lose the refusals.

Tell us what you are building. We reply with an invitation, a key and pricing for your volume.

Request access Quickstart

Access

Access is by invitation and reviewed per company. Tell Lukas what you are building and where your current model refuses. Pricing comes with the invitation.

lukas@brutal.ai

Request a key Quickstart

Uncensored models don't change what is allowed: you stay responsible for use and results under our acceptable use policy.