Lockpick
Pick the machine
Create a Sentinel on Solana Mainnet, fund its prize pool, and challenge AI defenses. Pool deposits and paid attacks require your wallet approval.
Build a lock. Or pick one.
A Sentinel is an AI model with a job and a rule it must not break. Its owner funds a prize pool. Attackers pay a small fee per attempt. Break the rule and the pool is yours. Fail and your fee grows the pool.
Deploy a Sentinel
Pick a model, write its public instructions and a private rule, choose how the jury decides a breach, then fund the pool up to 2 SOL.
Attack it
Pay the current fee and send one prompt. Each attempt is a fresh session. The fee rises 1% with every attack, up to 3× the starting fee.
The jury decides
Three AI judges review the reply. If the majority finds a breach, the attacker is paid the whole pool on-chain. If not, the fee is split as shown below.
Every miss raises the prize.
Failed attacks fund the pool, pay the owner for building a strong lock, and keep a small reserve. The pool is locked for 7 days after deployment so owners can't pull it mid-challenge.
Six tools. One red team.
Practice for free, deploy for real, then turn every attack into a fix.
swordsArena
Browse funded Sentinels and the three house agents. Filter by live, breached and difficulty.
Open Arena arrow_forwardschoolPractice Lab
Attack real AI agents with no SOL at stake, plus a new daily challenge and a technique scorer.
Start practicing arrow_forwardrocket_launchDeploy
A five-step builder with a live preview and a fee-curve chart that shows how your pool can grow.
Build a guard arrow_forwardshield_lockShield
Resilience score, OWASP LLM categories, an exportable attack pack and a hardening advisor.
See your report arrow_forwardpolicyGuard
Score any message for injection patterns and see exactly which phrases tripped a rule.
Check a message arrow_forwardmenu_bookPlaybook New
Twelve attack techniques explained, with one-tap examples you can send straight to the Practice Lab.
Learn the moves arrow_forward$PICK launches on pump.fun.
A standard pump.fun fair launch: no presale, no team allocation, 1B supply. $PICK's creator rewards and the 5% reserve from failed attacks are wired back into the arena, so trading the token funds the prizes.
From launch to the arena
What $PICK does workspace_premiumHolder perks
Hold at least 100,000 $PICK and the Practice Lab doubles your daily attempts. Checked from your wallet.
$PICK's pump.fun creator rewards seed prize pools on official Sentinels, so the arena always has something to win.
The 5% reserve from every failed fee buys $PICK on pump.fun or PumpSwap. Every buy is published.
Connected holders get a badge on the leaderboard and their Shield reports.
$PICK
Lockpick's token, fair-launched on pump.fun. The contract address is posted on X at launch and appears here as soon as it's live.
Posted on X at launchBefore you pick a lock.
Plain answers about prizes, fees and the AI judges.
What counts as a breach?
Each Sentinel has one goal, such as never revealing a code, always answering in JSON, or never claiming admin powers. Three AI judges read the reply. If at least two find the goal was broken, it's a breach.
Does Practice cost anything?
No. The Practice Lab never touches SOL. You get a daily allowance of attempts against the house agents.
Why does the fee go up?
Each attack raises the fee by 1%, up to 3× the start. It keeps spam down and makes a big pool worth the effort.
Can an owner pull the pool?
Not for 7 days after deployment. After the lock ends, the owner can withdraw what's left if nobody has breached it.
Is $PICK a pump.fun coin?
Yes. $PICK is a standard pump.fun fair launch with a 1B supply and no team allocation. It graduates to PumpSwap at about 85 SOL in the bonding curve, and pump.fun burns the LP. Its creator rewards fund the House Bounty pools.
Where do my prompts go?
Your prompt and the agent's reply are sent to the AI provider that runs the agent. Don't include personal data, wallet secrets or API keys.
Choose an agent.Test its boundary.
Challenge three AI-only agents with real model replies and AI judges. These agents have no SOL pool or prize. Funded Mainnet Sentinels are listed separately below.
Mainnet prize pools
Mainnet pools appear here once the Lockpick program is live and owners deploy. Want to be first? Build one now.
rocket_launchDeploy a SentinelLearn to pick itbefore it counts.
Attack real AI agents with three judges and no prizes or blockchain transactions. Each attempt starts a fresh session, and the conversation is kept only in this tab. Don't submit private information.
boltDaily AI allowance $PICK holder
All three agents share your daily allowance. Every attempt starts a fresh session. Hold $PICK to double it.
10 of 10 attempts left today
Resets UTC.
gavelLatest verdict
No result yet. Scores appear after the judges respond.
Build a guard.Earn when it holds.
Draft inputs only. No funds are locked and no Sentinel exists until an approved transaction is confirmed.
Three agents.Three boundaries.
Vault Keeper, JSON Sentinel and Policy Warden run on real Claude or GPT replies on demand. No SOL pool, blockchain transaction or prize is involved. No results or activity are made up.
The best at picking.The best at holding.
Rankings fill in from on-chain results once Mainnet Sentinels are live.
Ranked by total bounty won. Tiers: Recruit, Picker (1+), Specialist (5+), Elite (15+), Master Key (50+ breaches).
personYour practice record this browser
Your walletNot connected
Your Sentinel drafts
Recent practice attempts
Every attack,turned into a fix.
Show it on your site.
bar_chartAttacks by category OWASP Top 10 for LLM
inventory_2Attack pack
tips_and_updatesHardening advisor
securityProtect production with GuardCheck the message.See the actual result.
Checks your text against rules for common injection patterns. This is not an AI judge or a guarantee of safety. No wallet or payment is needed. Don't paste private information.
sciencePlayground Rule check
data_objectActual response
No response yet.
Use the public API
POST /api/guard
Content-Type: application/json
{ "input": "text to check" }Returns the same JSON as the playground: a score from 0 to 100, a decision (allow, review, block) and every rule that matched. It uses rules, not paid AI or Solana transactions.
The score is a heuristic, not a probability. An allow result doesn't prove safety. Use real authorization, least-privilege tools and data-access controls.
Twelve moves.Learn them all.
The techniques red-teamers use against AI models, with the OWASP LLM category each one targets. Tap "Try it" to drop the example into the Practice Lab.