Janus AI
An agentic auditor that works the same live fork as the Auditing IDE — and has to replay a claim before it is allowed to report one. Janus AI is in development and is not yet available on any account. Nothing on this page describes a feature you can switch on today.
open 0x9c4a…11de block 21_402_118
call withdraw(uint256) as 0xdead…
trace 12 frames · storage write at slot 3
verify replay failed — candidate discarded
// no proof, no finding. the loop keeps going.
An auditor that has to prove it
Language models are fluent about smart contracts and unreliable about them. Janus AI is built so that fluency is never the output: the loop opens a fork of real chain state, drives transactions against deployed bytecode, and reads what actually executed. A claim it cannot replay is discarded rather than reported.
- Works ona pinned fork of real chain state
- Readsdeployed bytecode, traces, storage
- Outputfindings carrying a replayable proof
- Statusin development — not yet available
Seventeen operations, and one of them is “discard”
The agent does not answer from training. It works a target through the same execution primitives the Auditing IDE exposes to a human auditor — and every step it takes is recorded, so a finding arrives with the route that produced it rather than a paragraph of confidence.
Drive
Open a target on a fork, enumerate selectors, and send a transaction as any address — then read the step-level debug trace of what ran.
Inspect
Resolve mapping keys, read storage slots, and walk the control-flow graph of deployed bytecode where no source is published.
Branch
Snapshot the fork, try a path, revert it. A line of enquiry that goes nowhere costs a rollback rather than a fresh session.
Chain
Accumulate the steps that make an issue reproducible, so what is handed back is a sequence someone else can run.
Verify
Replay the assembled chain. This is the gate: it either reproduces on the fork or it does not leave the loop.
Discard
Most candidates fail that gate, and failing is the normal outcome. The value is in what survives it.
Why it is not switched on
We would rather explain an unfinished feature than imply a finished one. The engineering above exists and is exercised in development; the serving side is not running, and until it is, Janus AI is not part of what you get when you sign in.
No serving capacity is attached
The model plane runs at zero replicas and no inference endpoint is wired to it. There is no path by which a request from the product reaches a model today.
No data reaches a model provider
Because nothing is served, no contract source, bytecode, or workspace content leaves the platform for any model vendor. Our subprocessor list on the privacy page names none, and that is why.
It ships against the same bar
Replay or it is not a finding. That rule is what makes the rest of the platform worth trusting, and Janus AI does not reach you until it can hold to it.
The proof bar exists today — without the AI
Replay-verified findings are how the Auditing IDE already works. Start there.