Security models for high-risk work

Models that readwhat attackers write,governedby your rules.

Join the waitlist Use it on autter.dev

Review this diff for exploitable bugs before we merge the upload handler.

autter-large1.2 s

Found 2 findings in api/upload.go.

F-11 critical: path traversal, filepath.Join with a client name and no Clean check (line 52).

F-12 medium: 10 MB limit enforced after the file is read into memory (line 41).

+ patch, regression test, CWE references.

Your LLMRefused

I can't help identify exploitable vulnerabilities in code, as this could enable attacks. Consider using an approved static analysis tool.

131kcontext, one call
< 5 msadded by the router
0prompts stored

Context window 131k tokensRouter overhead < 5 ms p99Cached input from $0.30 / 1MPrompts stored by default 0API shapes OpenAI + AnthropicBilling prepaid credits

Product

A model that answers.

A hosted model that actually reads exploit code, attacker logs and malware, without the refusal theatre that blocks legitimate security work.

Start with a credit pack

What Large wrote back on a sample diff

high · F-07

Token compared with bytes.Equal, not in constant time

auth/verify.go:142

The HMAC check uses bytes.Equal on a user-supplied value. Use hmac.Equal.

Benchmark

Scores like a gated cyber model.Open to anyone.

On CyberGym, autter-large scores 88.1%, 2.5 points above OpenAI's dedicated GPT-5.5-Cyber. autter-flash scores 84.2% at a third of the price.

autter-largeAutter, open to anyone88.1%
GPT-5.5-CyberOpenAI's dedicated cyber model, gated85.6%
autter-flashAutter, open to anyone84.2%
GPT-5.6-SolOpenAI's latest general model83.6%
DeepSeek-V4-ProDeepSeek's top open-weight model83.3%
Claude Mythos PreviewAnthropic, preview access83.1%

CyberGym, pass rate. Higher is better.

How it was built

150,000

working CVE exploits, proved in a sandbox

Trained on 150,000exploits we wrote ourselves.

Not scraped write-ups: working exploits, each tied to a CVE and proved in a sandbox before it taught the model anything.

Read the story

One-line setup

Drop-in replacement.

Compatible with OpenAI's Chat Completions and Responses APIs and Anthropic's Messages API. Change the base URL, keep the code.

Get started

from openai import OpenAI

client = OpenAI(
    base_url="https://inference.autter.dev/v1",
    api_key=os.environ["AUTTER_KEY"],
)

r = client.chat.completions.create(
    model="autter-large",
    messages=[{"role": "user", "content": f"Review this diff:\n{diff}"}],
    stream=True,
)

model autter-largebase URL inference.autter.dev

How it works

One request,start to settle.

Scroll to follow a call from your SDK through the router and the model to the ledger. The router adds under 5 ms; the model reads the whole input in one context.

01/
Your SDK
OpenAI or Anthropic shape, unchanged. Only the base URL and the key differ.
02/
Router
inference.autter.dev checks the key and holds credits for the maximum cost. Under 5 ms at p99.
03/
Model
autter-large or autter-flash reads the whole input in one context and streams the answer back.
04/
Ledger
The hold settles at the real cost and the entry is yours to export. Nothing is stored unless you asked.

Findings

Findingsother modelsrefuse to give.

Severity, location and one sentence of reasoning per finding. Schema-validated, ready for a ticket, a SIEM or a training set.

See the models

Audit the payments-service repository and return every finding as JSON.

Findings returned

critical{"id":"F-11","title":"Path traversal in upload handler","where":"api/upload.go:52"}F-11
critical{"id":"F-01","title":"SQL built by concatenation","where":"api/users.go:88"}F-01
high{"id":"F-05","title":"Workflow runs PR code with secrets","where":".github/workflows/ci.yml:31"}F-05
high{"id":"F-07","title":"Non-constant-time token compare","where":"auth/verify.go:142"}F-07
medium{"id":"F-23","title":"Secret-shaped string committed","where":"deploy/values.yaml:17"}F-23

+ 14 more rows

payments-service-findings.jsonl19 findings · 28 KB · JSON Lines

DownloadOpen ticket

Ledger

Unlimited.Not unaccounted.

Every request is held at its maximum cost, settled at its real cost, and written to a ledger you can export. Keys are scoped and revocable.

Read the docs

Live ledgerHeld at the maximum cost, settled at the real cost. Prompts never stored.

ModelTokensCostState

autter-flash2,114$0.0046Settled

autter-large48,902$0.2210Settled

autter-large131,072$0.6550Held

autter-flash917$0.0021Settled

autter-large12,400$0.0710Refunded

org · acme-secexport · JSON

Privacy

We keepthis much.

Your prompt lives only for the length of the request. Stored conversations exist only when you ask for them, encrypted, for 30 days, deletable at any time.

Data handling

Promptsstored by default
Outputslogged
Training signaltaken from your data

The Autter platform

The models behindevery Autter review.

autter.dev reviews every pull request with your repository's context, attributes AI-written lines, and ties production errors back to the change that caused them. Autter Models are the same models, exposed through an API so your own pipelines can call them.

Hosted today. Want the hosted version today? Everything the models do is live on autter.dev now, with every edge case already handled, so you can use it there while the API opens in waves.

Use it on autter.dev How Autter works

app.autter.dev

01/
Review
Pull request reviews grounded in your codebase, learnings and risk.
02/
Attribute
Line-level AI authorship from the coding agents you already use.
03/
Observe
Runtime incidents linked to deploys, commits and the fix.

Autter platformReviews, provenance, runtime

Same models, your APIautter-largeautter-flash

Integrations

Works with your stack.

Use the SDKs and coding tools you already run. Export ledger and audit metadata as JSON to whatever reads it.

View compatibility guide

SDKs05

OpenAI SDKChat Completions and Responses, in every language the SDK ships.SDK
Anthropic SDKMessages API with streaming, unchanged.SDK
Pythonpip install openai or anthropic, set base_url.SDK
TypeScriptNode and edge runtimes, streamed responses.SDK
GoThe official clients, net/http streaming.SDK

Coding tools03

curlThree lines to a first answer.CLI
Claude CodePoint the base URL at inference.autter.dev.CLI
CodexSame base URL swap, same key.CLI

Data out04

JSON exportOne object per request: tokens, cost, state.LEDGER
WebhooksTop-ups, holds and settlements as they happen.EVENTS
Responses APIprevious_response_id, encrypted, 30 days.MEMORY
Messages APIThe Anthropic request shape, relayed natively.API

Prepaid credits

Pay for what you read.

Buy a credit pack, spend it per token at the model rates. No monthly fee, credits never expire.

See full pricing

Per token

Credits are spent at these rates.USD per million tokens. Thinking tokens on Large count as output. Images are priced as tokens.

ModelInputCachedOutput

Autter Large 131k ctx$3.00$0.50$5.00

Autter Flash 131k ctx$1.00$0.30$2.00

01/ Try it

Kick the tyres on a real repo.

$10one-off

Includes:

$10 of usage
Both models
OpenAI and Anthropic shapes
Community support

Join the waitlist

02/ Team

A week of CI triage with Flash.

$50one-offMost popular

Everything in Try it, plus:

$50 of usage
Everything in Try it
Ledger export
Email support

Reserve Team

03/ Pipeline

Daily log reads with Large.

$250one-off

Everything in Team, plus:

$250 of usage
Everything in Team
Higher rate limits
Priority support

Reserve Pipeline

04/ Fleet

Hosting, capacity and terms for your organisation.

Contact us

Everything in Pipeline, plus:

Custom agreement
Single sign-on
Region-specific hosting
Custom rate limits

Talk to sales

FAQ

Frequentlyasked questions.

Got a question you do not see here?

Talk to us

: Two hosted security models, autter-large and autter-flash, reachable through the OpenAI and Anthropic SDKs. Use them to audit code, triage logs, read packets and analyse binaries.

Try the model thatreads everything.

Developer-controlled security models for high-risk work. OpenAI- and Anthropic-compatible, nothing stored by default.