Guardial - Mission-driven AI red team. Real proof. Zero guesswork.

What inspired us to build Guardial? We kept seeing the same failure mode in AppSec. Scanners fire fixed payloads and dump hundreds of “findings.” Most are noise. Manual red teams do the real work recon, hypothesize, attack, watch the response, adapt, chain impact but that doesn’t scale, and it doesn’t run every week. Then AI entered the picture. Most “AI security” tools wrap an LLM around a scanner and let the model decide what’s vulnerable. That scared us. An LLM can invent a convincing story. A bank can’t ship remediation on a hallucination. So we asked a different question: what if the AI acted like a red teamer’s brain planning and adapting but never got to declare a finding confirmed? That’s Guardial. The doctrine is literally in our codebase (confirmation_invariant.go): MODEL PROPOSES. CODE EXECUTES. CODE OBSERVES. CODE PROVES. LLM NEVER BECOMES THE SECURITY ORACLE. How it evolved: • We started with traditional OWASP/detection tooling. • We realized AppSec needed the red-team loop recon → plan → attack → observe → replan → chain not another checklist. • We built an autonomous engine: an LLM Brain selects specialist workers (injection, IDOR/BOLA, XSS, SSRF, auth, secrets…), workers run real HTTP/browser attacks, and multi-signal oracles confirm negative controls, OAST callbacks, browser dialogs, timing diffs, live two-user owner-bound replay for BOLA. • A finding can only be stored as “confirmed” when bound deterministic proof exists enforced at the database layer. The LLM can demote a finding; it can never upgrade one to confirmed. • Kill chains and an attack graph stitch proven hops into impact, with curl PoCs, SARIF/HTML/PDF, and retest recipes. What problem we’re solving: Security teams don’t need more alerts. They need offensive proof “can another user access this account?” answered with a reproducible attack, not a CVSS guess. Especially on APIs, where BOLA/IDOR is the #1 risk and scanners routinely miss it. Where we’re going: Mission-driven red teaming type an objective (“prove IDOR on /api/accounts/{id}”, “hunt leaked secrets”) and Guardial plans, executes, and returns a verdict with evidence. Same prove-first rules. Same RoE-safe, non-destructive defaults. We’re not building another VAPT report factory. We’re building an autonomous red team that earns trust the hard way: proof before confidence. Would love feedback from AppSec, red teamers, and founders shipping APIs what’s the one objective you’d type first?