Skip to content
CyrailsAI

Site menu

Abliterated models. Security research.

No guardrails · Offensive security

Security research,with less friction.

Abliterated models for authorized offensive security work - trace trust boundaries, draft PoCs, and correlate logs with 1M-token context.

For pentesters, bug bounty hunters, and security analysts

Authorized scope only. Review data handling before upload.

Model rosterone endpointGLM 5.3abliteratedGLM 5.3 FlashabliteratedDeepSeek V4.1 Flashabliterated1Mtoken context
Use cases

Built for the work
behind the finding.

One focused workspace for the evidence-heavy parts of security research.

01 / Use case

Bug bounty

Trace trust boundaries, inspect related code, and draft PoCs for validation.

auth.middleware.ts
02 / Use case

Threat hunting

Correlate sanitized logs, compare event patterns, and develop investigation hypotheses.

events.ndjson
03 / Use case

Malware analysis

Explain obfuscated scripts and inspect payload behavior from code you supply.

payload.js
The models

Specs up front.

Choose by context, output limits, and listed throughput.

Z.aiDeep trace

GLM 5.3

Abliterated

Trace trust boundaries and work through layered vulnerability hypotheses.

Context
1M tokens
Max output
131,072 tokens
Listed speed
85 tok/s
Quantization
NVFP4
Endpoint
/v1/chat/completions
glm-5.3-abliterated
Z.aiFast iteration

GLM 5.3 Flash

Abliterated

Iterate on test ideas, inspect code, and explain unfamiliar behavior.

Context
1M tokens
Max output
128,000 tokens
Listed speed
150 tok/s
Quantization
NVFP4
Endpoint
/v1/chat/completions
glm-5.3-flash-abliterated
DeepSeekResearch mode

DeepSeek V4.1 Flash

Abliterated

Analyze source, draft research scripts, and review proposed fixes.

Context
1M tokens
Max output
131,072 tokens
Listed speed
150 tok/s
Quantization
FP8
Endpoint
/v1/chat/completions
deepseek-v4.1-flash-abliterated

What changes: Abliteration targets refusal behavior; zero refusals are not guaranteed.

What stays: Base-model limitations remain. Validate reasoning, generated code, and findings before acting on them.

Listed speeds are supplied specifications, not independent benchmarks. CyberGym / DEFCON scores are not published.

Workspace + API

Same models.
Your tools.

Start in chat. Build into your research workflow.

01 / The workspace

Keep the evidence together.

Bring repository files, extracted PDFs, and sanitized logs into one conversation. Compare more evidence with a 1M-token context window.

02 / The API

Call it from your own code.

Use the same models through an OpenAI-compatible endpoint in the tools you already work in.

Long-context research
1M tokens

More context. Not unlimited storage.

Fit depends on tokenized input and output needs, not automatic GitHub ingestion.

research.pyPython · OpenAI SDK
from openai import OpenAIclient = OpenAI(    base_url="https://api.cyrails.com/v1",    api_key="YOUR_CYRAILS_API_KEY")result = client.chat.completions.create(    model="glm-5.3-abliterated",    messages=[        {            "role": "user",            "content": "Review this code for access-control flaws."        }    ])print(result.choices[0].message.content)

A larger context window does not guarantee exhaustive analysis of every file. Confirm data handling before submitting proprietary vulnerabilities.

Your evidence
stays yours.

Confirm data handling before submitting proprietary vulnerabilities. This module records what is known and what still needs to be verified with the team.

Third-party inferenceZero retention · not establishedSOC 2 · unverified
Confirm current terms with the team
Data handling ledger
  • Workspace data

    Saved conversations support continuing investigations across sessions.

    saved
  • API processing

    Confirm prompt logging, retention, and inference-provider handling with the team.

    confirm
  • Assurance

    SOC 2 status is unverified. Zero data retention is not established.

    unverified

Third-party inference providers are used. A zero-retention guarantee is not implied; confirm current terms before uploading sensitive material.

Pricing

Pay as you go.

Prepay credits and spend them across chat and API. No monthly subscription.

Pay as you goPrepaid credits · one balance

Top up once.
Spend it as you work.

Buy credits when you need them. One balance covers workspace chat and OpenAI-compatible API calls. Credits are not tokens; model credit rates are shown in your workspace before a request runs.

  • Prepaid credits, spent as you work
  • Chat and API share one balance
  • Model credit rates shown in your workspace
  • Top up only when the balance runs low
Credit balancechat + api

One balance

Metered at the model rate shown in your workspace. Top up at any time.

Workspace chat
metered
API calls
metered
Monthly fee
none
Shared wallet

Workspace chat

Threads, saved conversations, and research notes draw from the same balance.

/v1

OpenAI-compatible API

Every /v1/chat/completions call is metered at the selected model's credit rate.

By agreement

Enterprise

Committed capacity, invoicing, and access terms agreed before activation.

Credits are not tokens. Model rates apply and are shown in your workspace. No unlimited-token allowance is implied. An API free tier has not been announced. Enterprise capacity and invoicing are agreed before activation.

F.A.Q

Direct answers.

Ask the curated assistant about refusals, context, data handling, pricing, or API access. Answers come from the product FAQ.

Curated answers. Not a live model.

Cyrails assistantFAQ

Curated answers

CyrailsAI

What do you want to know about Cyrails?

Curated answers on refusals, models, context, data handling, pricing, and API access.

Enter to send · Shift+Enter for new line

Cyrails assistant

FAQ

Curated product answers

CyrailsAI

What do you want to know about Cyrails?

Curated answers on refusals, models, context, data handling, pricing, and API access.

Enter to send · Shift+Enter for new line

Curated answers from the product FAQ, not a live model. Confirm current terms before submitting sensitive data.

Get started

Bring the question.
Inspect the output.

Start in the workspace, then call the same models from your own tooling. One credit balance across chat and API.

3 models · 1M context
cyrails / readylive

$inspect cyrails --models

GLM 5.31M

GLM 5.3 Flash1M

DeepSeek V4.1 Flash1M

›Bring the question.

OpenAI-compatible