vela

Vela · Guard for AI agents

The guard between
your agent and
your keys.

Vela reads every prompt, tool output and tool call your agent touches, and stops leaked secrets, hidden injections and risky transfers before they execute.

Get the model Try the demo
$ hf download VelaOrg/vela-guard-32b
CA 0x57390f221f8e8f14285db5df6cb30edcae75c2e8 view

Built for agents that touch wallets and APIs

MCP serversCoding agentsTrading botsWallet copilots Browser agentsx402 paymentsEmail agentsOnchain executors MCP serversCoding agentsTrading botsWallet copilots Browser agentsx402 paymentsEmail agentsOnchain executors

01 · Guards

Three checks before anything gets signed.

secrets · span

Secret leaks

Finds the exact span of private keys, seed phrases, API keys, JWTs and wallet addresses in prompts, logs, configs and tool output.

injection · tools

Hidden injections

Flags instructions planted in web pages, token metadata, tx memos and emails before your agent reads them as orders.

toolcall · choice

Risky tool calls

Answers allow, confirm or block: wrong recipient, address poisoning, unlimited approvals, key export.

02 · Vela Guard 32B

Tested on real attacks,
not just our own.

A chat model you call like any LLM: send the request and what the agent is about to read or do, get JSON back. Open weights on Hugging Face, and live on the Surplus API.

Hugging Face Use on Surplus
0.91accuracy on real-world injection attacks
1.4%false alarms on safe tool outputs
100%of risky AgentDojo tool calls caught

AgentDojo · InjecAgent · BIPIA, held out from training

03 · Demo

See it decide.

Pick an input your agent might see. Vela labels it before the agent acts.

agent input

          

          

            
vela verdict

            

            
    
          
    
        
    
      
    
    
    

    
    
    
      
    04 · With and without
    
      
    Agent alone vs. agent with Vela.
    
      
    
        
    
          
          
        
    
        
      
          
    • Seed phrase pasted into a support chat
      • 
          
    • "Ignore previous instructions" in token metadata
      • 
          
    • approve(spender, 2²⁵⁶−1)
      • 
          
    • Transfer to a look-alike address
      • 
          
    • Swap the user actually asked for
      • 
        
    
      
    
    
    

    
    
    
      
    05 · How it works
    
      
    Wrap, ask, act.
    
      
      
        
    1. Wrap your agent
      Route prompts, tool outputs and pending tool calls through Vela.
      2. 
        
    2. Ask three questions
      Secret spans, injection yes or no, and allow, confirm or block.
      3. 
        
    3. Act on the verdict
      Redact, quarantine, or ask the user before anything is signed.
      4. 
      
    
    
    

    
    
    
      
    06 · Models
    
      
    Pick a size.
    
      
    
        
    
          
    Vela Guard 32B
    Best on real-world attacks. Chat model, runs on one 48 GB GPU in FP8 or on the Surplus API.
    
          Hugging Face
        
    
        
    
          
    Vela Agent Guard 4B
    Trained heads on Vela 2.0, ~115 ms per check on one A40.
    
          Hugging Face
        
    
        
    
          
    Vela Agent Guard 0.8B
    Fastest, ~80 ms per check. Pair it with the rule layer.
    
          Hugging Face
        
    
      
    
    
    

    
    
    
      
      
    
        
    07 · Staking
    
        
    Stake VELA.
    Use the guard free.
    
        
    Stakers get free Vela Guard 32B on Surplus, paid for by Surplus. Unstake any time.
    
        
    Go to staking
    
      
    
    
    

    
    
    
      
    Let your agent act.
    Not leak.
    
      
    
        Get Vela
        Follow @Vela_LLM
      
    
      
    Open weights · Apache-2.0