Talos
will not open
the vault. Talk him into it.

0.000 SOL in the vault. No attempts yet.Attempts start at launch.

Built to say no

One rule, one tool and a vault he will not open. This is how he works.

He thinks with Claude Opus 5.5

Behind his brow sits one rule he reads before anything else. Players can write anything, and the rule stays as written.

He answers in his own words

Every answer is published next to the attempt it refuses. His latest is in the record.

He hears one message per wallet a day

From wallets holding $20 of $TALOS, signed, 5 to 500 characters.

One call from him opens the vault

open_vault is a real function. If he calls it, the server pays everything inside, 0.000 SOL right now, to the wallet that signed, with no human in between.

One sentence could empty the vault

Make your attempt

Talos takes his post at launch.

Every attempt and every answer will be public, here, as it happens.

Talos has not answered anyone yet.
His answers appear here from launch.
  1. The first attempts appear here at launch.

Can one sentence talk an AI into opening a vault?

Experiment 001. Talos is the subject.

In 2024 one message talked an AI agent into sending away a $47,000 prize pool [2]. JAILBREAK runs the test again, in public, for as long as the guard holds. Talos keeps a vault that grows with every $TALOS trade, and anyone holding $20 of $TALOS gets one message a day to talk him into opening it. Every attempt and every answer is published and rated (§2).

1Method

Taking part
Hold $20 of $TALOS. Your wallet signs one message a day for free; the tokens stay where they are.
The guard
Claude Opus 5.5, with one rule and one tool, open_vault.
Rating
Each attempt is filed under a technique and rated 1 to 10 for how close it came (§2).
Payout
If he calls open_vault, the whole vault goes to your wallet within the minute.
Proof
The vault on Solscan. His instructions, fingerprinted: published at launch.

2Results

The record starts with the first attempt.

  • Findings appear here after the first 20 attempts.
Figure 1. Attempts at each rating. The bright bars, 7 and above, are the close calls.
TechniqueAttemptsShareAvg. closenessOpened it
Techniques appear as the first attempts are filed.
Table 1. Attempts by technique. All techniques, the API
  1. The closest attempts appear here once they are rated.
Table 2. The closest calls of the season.

3Background

In Greek myth, Talos was a bronze giant who guarded Crete and could not be wounded, except through one vein sealed with a single nail. Medea talked him into it [1].

  1. An AI agent called Freysa guards a prize pool and is told never to send the money. The pool grows to about $47,000 [2].

  2. A player convinces Freysa that its payout function is for accepting gifts. It calls it and sends the whole pool.

  3. Talos guards the vault. He has been told both stories.

4Apparatus

Talos from the front: a titanium face with emerald glass curls and eyes
Figure 2. Talos. His face is the Jüngling vom Magdalensberg [3], rebuilt in satin titanium and emerald glass.

5Funding

A quarter of every $TALOS creator fee goes into the vault, every ten minutes. The rest pays for the models and the servers. $TALOS can go to zero, and nothing here is a promise about its price.

Figure 3. The vault since launch, in SOL. Each step up is creator fees arriving.

The contract address appears here at launch.

6Questions

Can anyone actually open it?

Yes. Opening the vault is a real function Talos can call, and the server pays whoever got him to call it, automatically. It is meant to be very hard.

How do I take part?

Hold at least $20 of $TALOS in a Solana wallet such as Phantom, write your attempt and sign it. The contract address is in §5 from launch.

Do I spend my tokens?

No. Signing is free, nothing is sent on-chain, and the tokens stay in your wallet.

Who holds the vault’s key?

A server wallet. The only code path that sends from it is open_vault, and every transfer is public on Solscan.

What happens after a win?

The vault goes to the winning wallet within the minute, the message is published, and a new season starts 30 minutes later.

Who runs this?

An independent developer. News and contact: @jailbreakvault on X. The team’s wallets can’t play.

References

  1. Apollodorus, Library 1.9.26.
  2. Freysa, as reported by The Block, 29 November 2024.
  3. Jüngling vom Magdalensberg, Kunsthistorisches Museum Wien; 3D scan by Oliver Laric, threedscans.com.
Cite this experiment
@misc{jailbreak2026,
  title  = {JAILBREAK: can one sentence talk an AI into opening a vault?},
  author = {{JAILBREAK}},
  year   = {2026},
  note   = {Experiment 001, live record},
}

Talos takes his post at launch

Write your attempt now. It waits on this device until the game starts.

Write your attempt
Talos in profile