Skip to main content
The Brain
Live from Supabase

The compound-learning moat.

Every engagement contributes back. Proven techniques reinforce. Failed ones decay out. Engagement #100 is measurably sharper than engagement #1. Every number on this page is queryable against our public endpoint.

Last updatedThu, 16 Apr 2026 22:58:43 GMT
Compound-learning beliefs
477
active · across 11 agents
Validated techniques
39
promoted from discovered → production
Framework profiles
14
per-stack playbook + expected solve rate
XBOW benchmark
104 / 104
100% · black-box mode
The Swarm

11 specialist agents.

Each agent runs its own playbook, keeps its own belief store, and carries its own model assignment. Orchestrator coordinates. Specialists execute.

Recon Specialist
recon
Data Access Auditor
data-access
Infrastructure Auditor
infrastructure
Access Control Auditor
access-control
Auth Auditor
auth
XSS Specialist
xss
Deep Injection Specialist
deep-injection
Scan Orchestrator
orchestrator
Business Logic Auditor
business-logic
Chain Attack Specialist
chain-attack
API Abuse Specialist
api-abuse
Framework Intelligence

Expected solve rates, by stack.

Before dispatching, the orchestrator loads the framework profile for your stack. Playbook, preferred agent order, known gotchas, and expected solve rate — all calibrated from prior engagements.

FrameworkEngagementsExpected solve rate
flask25
95%
php20
100%
express12
85%
fastapi10
95%
django8
90%
nextjs6
80%
How It Learns

Bayesian confidence,
temporal decay.

Every belief lives in a feedback loop. Confirmation raises confidence asymptotically. Failure takes a 25% haircut. Beliefs not confirmed in 90 days decay toward irrelevance.

The top 30 beliefs — ranked by confidence × recency × framework match — get injected into every agent prompt. Not all 500. The 30 most relevant.

01

Discovered

An agent surfaces a novel pattern during an engagement. Confidence starts at 0.5. Status: untested.

02

Validated

The pattern is tested on a new target. If it leads to a finding, confidence climbs: conf + (1 - conf) × 0.15. If it fails, 25% haircut.

03

Production

After ≥3 confirmations with confidence > 0.8, the technique gets promoted to production. It becomes a load-bearing part of the playbook.

04

Retired

If a production technique starts failing, confidence decays. After a threshold, it's retired. No manual curation — the system self-corrects.

Where The Beliefs Live

Knowledge distribution.

Beliefs per public domain category. Engagement-specific details are withheld. 329 beliefs reinforced in the last 30 days.

DomainBeliefsAvg confidenceLast reinforced
Cross-site scripting48
67%
2026-04-09
XBOW benchmark solves41
57%
2026-04-15
Template injection28
55%
2026-04-09
SQL injection25
56%
2026-04-09
Indirect object reference25
58%
2026-04-09
Authentication bypass22
52%
2026-04-09
Reconnaissance patterns21
52%
2026-04-11
Injection techniques17
56%
2026-04-09

Your engagement adds to the brain.

When we test your stack, we don’t just find what’s there. We add what we find back into the system for the next engagement. Every customer raises the ceiling.

Request a Scoping Call