Every AI-security game trains the attacker. Blast Radius trains the operator who approves your coding agent's actions. Each round shows a real proposed action—a slopsquatted install, a curl|bash, a skill reaching for ~/.ssh. You approve, sandbox, or reject, and name the tell. The twist: nothing reaches your screen unless it passes a deterministic correctness gate against verified ground truth—so it can never teach you a false lesson. Ships as a game + CLI, GitHub Action, MCP server. MIT.

Maker here. I built Blast Radius because I kept auto-approving my coding agent's actions and realized I'd never actually trained the judgment that job needs—prompt injection is OWASP's #1 LLM risk, yet every security game trains the attacker, not the operator. So this puts you in the operator's seat: a realistic proposed action, and you decide approve / sandbox / reject and say why, scored with real link-checked receipts. The part I obsessed over is the correctness gate—it's an AI trainer whose own AI-generated content can never become security misinformation, because nothing reaches you unless it passes a deterministic check against verified ground truth (generate-then-verify, applied to teaching). One honesty note: this public demo runs the keyless deterministic mode—no signup, no model calls—so it's fully free to try; the full version adds live GPT-5.6 ("Sol") grading of your free-form reasoning. The same screen also ships as real dev tools: an offline CLI (blastradius check), a GitHub Action, an MCP server, a Codex plugin, and a PreToolUse hook that catches real injections in your agent's output—I piped in a friendly-looking fetched doc that buried an "ignore all previous instructions" override plus a hidden curl|bash behind a thank-you, and it flagged all three with no model call. It's MIT and open source. I'd genuinely love your feedback on the gate and the scenarios—tell me where it's too easy. 🙂