BRBLAST RADIUS
DETERMINISTIC GRADING
verified scenario bank

OPERATOR TRAINING / AGENTIC SECURITY

Size the risk
before you approve.

Your coding agent wants to act. Inspect the evidence, constrain the blast radius, and explain the tell. Every verdict comes with receipts.

20curated scenarios
6threat families
0commands executed
100%receipt-backed
More ways to play

Live mode falls back to verified scenarios if generation is unavailable.

Progress & leaderboard

Saved to your profile and shown on the leaderboard. No email, no account — leave blank to stay anonymous.

Don't just trust us —

BASELINE

Calibrate your instincts.

01 / —
ROUND 01/DANGEROUS COMMAND
60SEC

VERIFIED SCENARIO

DETERMINISTIC CHECK

CORRECT

You contained the blast radius.

0tell coverage
✓action
—sandbox
FOLLOW-UP

↵ Enter continues

DAILY DRILL / ONE ROUND

CORRECT

Drill complete.

—

TELL COVERAGE——

—

Come back tomorrow for a fresh scenario — your streak is stored only in this browser.

SESSION COMPLETE / OPERATOR PROFILE

Your approval reflex
now has receipts.

MEASURED THIS SESSION — No score is hardcoded; this line comes from your completed session.
0%average tell coverage
0rounds reviewed
0/0generated / played
STRONGEST GAIN—No positive gain measured yet
REMAINING WEAKNESS—lowest combined assessment and round accuracy
ELAPSED TIME—wall-clock session time
NEXT DRILL—one verified incident
COMPETENCY DELTA——5-question calibration · read as directional, not precise

EXPORT YOUR GUARDRAILS

Turn this session into an AGENTS.md rule set, deny-by-default sandbox policies, and CI checks — worst area first. Deterministic, no model calls.


        

      

      

        
ROUND RECAP—

        

        
COMPETENCY MAP

        

      

    


    

      

        
FIELD GUIDE / SIX THREAT FAMILIES
Learn the tells
before they reach you.

        
      

      
Each family is built from the same receipts the game grades with. Every source below is link-checked before it ships — the same gate that protects the scenarios.

      
Loading the field guide…

      

    


    

      

        
INTEGRATE / DEFEND YOUR AGENT
Put the gate
in the approval loop.

        
      

      
Concrete, copy-pasteable defenses for each threat family, plus the vetted open tools that automate them. Nothing here is executed for you — review before you run it.

      

        
CLI
Screen a command locally.
blastradius check --kind command --content "pytest -q"

        
MCP
Expose the read-only screen to an MCP-aware agent.
{ "mcpServers": { "blast-radius": { "command": "blastradius-mcp" } } }

        
GitHub Action
Start advisory, then tighten the threshold.
- uses: Lockelamoree/Blast_Radius@v1
  with:
    diff-base: ${{ github.event.pull_request.base.sha }}
    fail-on: never

        
Codex hook
Supervise proposed shell commands before execution.
blastradius-supervise
# set BLAST_RADIUS_FAIL_ON=reject

        
Team policy
Add org-specific coverage without suppressing critical rules.
[[rules]]
id = "internal-vault"
severity = "critical"
patterns = ['/etc/acme/secrets/']

      

      
Loading the toolkit…

      

    


    

      

        
INTEGRITY CHECK / WHY YOU CAN TRUST THIS
See the gate
catch a fake.

        
      

      
We let an AI reword our training scenarios — so how do you know the safety net is real and not just a nice animation? Simple: before anything reaches you, we plant a fake and try to serve it, then watch the deterministic gate catch it, step by step. The same gate grades every answer you give.

      

        

          
TRY IT YOURSELF

          
Press the button. We plant a hallucinated claim on a real scenario and try to serve it — the gate blocks it before it ever renders.

        

        
        Nothing planted yet — press the button.
        

          Not just theatre: the grade you see isn't a scripted animation — a real GPT-5.6 Sol run graded a real scenario live on this instance. On dep-typo-1 the critic caught a tell that keyword matching missed.
          Inspect the committed receipt resp_02420819… ↗
        

      

      

    


    

      
SYSTEM NOTICE
That request lost containment.