v0.4.2 — Google OAuth dashboard auth, 12 agents, 339 tests

Your codebase's
immune system

An autonomous platform that scans, predicts, quarantines, and fixes bugs across your entire estate — then learns from what worked.

Book a demo

Watch how BISO scans, predicts, and quarantines vulnerabilities — then learns from every fix.

Full Episode Podcast
Full Episode
BISO and the Bug Immune System
Founder deep-dive · Richard Leclézio · ~45 min
0:00~45:00
BISO — How the Bug Immune System works

Bug backlogs never shrink

Your team triages the same patterns month after month. Scanners find issues faster than humans can fix them. The backlog grows. MTTR drifts up. CISOs get nervous. Something has to change.

73%
of security debt is re-discovery
The same vulnerability class, in the same codebase, found again next quarter.
42d
average MTTR for high-severity bugs
From finding to merged fix. Most of that time is human context-switching.
growing backlog, fixed team
Scanners scale linearly. Your security team doesn't. The gap only widens.
Closed learning loop

Twelve agents. One loop. Always learning.

Every bug the system sees makes every future prediction sharper. That's the moat — not the code, the compounding memory.

Scan
Multi-scanner (Semgrep + CodeQL + Snyk) with cross-scanner dedup
Triage
Cluster, score, deduplicate — memory-aware fallback when offline
Predict
Oracle scores the change BEFORE merge — topology + memory + patterns
Quarantine
Contain the threat — flip feature flags or trip circuit breakers instantly
Fix
Adaptive blue → red → judge loop with iterate-on-failure, real worktrees
Remember
Historian embeds every bug into semantic memory — same pattern never re-triaged from scratch
Learn
Outcomes calibrate per-class + per-signal weights. The system sharpens.
every run makes the next one smarter

See it in action

Every screen from the live dashboard — populated with demo data in one click.

Live system status, agent cards, quarantine count, real-time event feed
Overview

Live system status, agent cards, quarantine count, real-time event feed

Escalated fixes under 'Awaiting your review', new bugs below — accept or retry in one click
Triage inbox

Escalated fixes under 'Awaiting your review', new bugs below — accept or retry in one click

Every fix loop the system has run — status, verdict, and duration at a glance
Fix loops

Every fix loop the system has run — status, verdict, and duration at a glance

Full timeline: worktree → blue diff → red test → runner → judge verdict
Fix replay

Full timeline: worktree → blue diff → red test → runner → judge verdict

Every PR scored with signals, verdict, and blast radius before it merges
Oracle

Every PR scored with signals, verdict, and blast radius before it merges

Force-directed import graph across 5 languages — click any file to explore
Cartographer

Force-directed import graph across 5 languages — click any file to explore

Flip a feature flag or trip a circuit breaker to contain a threat in seconds
Quarantine

Flip a feature flag or trip a circuit breaker to contain a threat in seconds

Adversarial input generation against a shadow-prod digital twin — finds bugs before users do
Fuzzer

Adversarial input generation against a shadow-prod digital twin — finds bugs before users do

Speculative vulnerability imagination — finds bugs static scanners can't see, overnight in parallel sandboxes
Dreamer

Speculative vulnerability imagination — finds bugs static scanners can't see, overnight in parallel sandboxes

Dollar value per bug — revenue at risk, SLA penalty, and eng-hours so you fix the expensive bugs first
Economist

Dollar value per bug — revenue at risk, SLA penalty, and eng-hours so you fix the expensive bugs first

Semantic memory — recognises returning patterns without re-triaging from scratch
Historian

Semantic memory — recognises returning patterns without re-triaging from scratch

Per-class + per-signal calibration from real fix outcomes — predictions sharpen over time
Learner

Per-class + per-signal calibration from real fix outcomes — predictions sharpen over time

Declare bug classes mathematically extinct — certificates auto-revoked if a counter-example appears
Extinction

Declare bug classes mathematically extinct — certificates auto-revoked if a counter-example appears

Cross-organisation bug pattern database — the 100-year institutional memory moat
Pathogen registry

Cross-organisation bug pattern database — the 100-year institutional memory moat

Scan rates, Oracle accuracy, MTTR trends, and agent activity over time
Metrics

Scan rates, Oracle accuracy, MTTR trends, and agent activity over time

Hash-chained immutable record of every agent decision — verifiable in one command
Audit log

Hash-chained immutable record of every agent decision — verifiable in one command

Kill switch, integrations status, storage backend, and system config
Settings

Kill switch, integrations status, storage backend, and system config

Scroll to see all screens — click any to expand

What makes BISO different

Not another scanner. Not another chatbot. A nervous system.

Topology-aware prediction

The Oracle doesn't just pattern-match the diff. It knows the import graph, the blast radius, the file's bug history, and the learner's calibration. A 1-line comment on a load-bearing file scores BLOCK. The same touch on a leaf? Safe.

Self-healing fix loop

Blue-agent proposes a minimal diff. Red-agent writes an adversarial test. Judge scores them. On failure, the loop iterates with the previous attempt as feedback. On escalation, the patch surfaces in the triage inbox — accept or retry in one click. Concurrent-fix lock prevents double-runs.

Semantic bug memory

Every bug is embedded into a 512-dim vector via deterministic hashed-TF. Triage recognizes returning patterns instantly. No API, no GPU, cross-process stable, works completely air-gapped.

Per-signal learner calibration

When a fix holds, regresses, or rolls back, the learner adjusts both per-bug-class and per-signal weights. Oracle's future predictions get sharper with every observed outcome.

Multi-scanner, cross-validated

Semgrep + CodeQL + Snyk running side-by-side. Findings two scanners agree on are marked cross_validated with boosted confidence. One orchestrator, three backends, zero duplicates.

Safety-first architecture

Default-deny policy engine. Hash-chained audit log. Kill switch that halts every agent in one command. Forbidden zones (auth/payments/IaC) are read-only. Every action authorized + logged.

GitHub PR auto-commenting

Connect a GitHub App and BISO posts Oracle verdicts directly on pull requests. Scores, signals, and verdict badges appear as PR comments and commit statuses — no context-switching required.

Quarantine — contain before fix

When a confirmed vulnerability hits a live code path, BISO flips a feature flag or trips a circuit breaker instantly. Contain the blast while the fix loop works — don't wait for a patch to stop the bleeding.

Triage inbox — incident-response UX

Escalated fixes surface under 'Awaiting your review' with an inline diff, test results, and judge rationale. Accept or retry in one click. New bugs queue below. No hunting through separate fix-loop pages.

Webhook alerts

Receive real-time notifications when Oracle blocks a PR, a quarantine is engaged, or a fix loop ships. Plug into Slack, PagerDuty, or any webhook consumer with zero configuration.

Rollout timeline

Enterprise integration

From zero to production in two weeks. Passive first, active second, steady state forever.

Day 1
2 hours setup
  • Docker compose up + GitHub App install
  • Index repos + baseline scan
  • Dashboard live with initial bug inventory
Week 1
Passive mode
  • Oracle comments on PRs — no blocking
  • Security Lead reviews signal quality
  • Historian builds semantic memory
Week 2
Active mode
  • Merge gating on — BLOCK = blocked
  • Slack + PagerDuty alerts enabled
  • Learner calibration begins
Steady State
15 min/day
  • Security Lead reviews triage inbox — accept escalated fixes in one click
  • Monthly audit + kill switch drill

26 agents. One orchestrator.

Each agent does one thing well. The orchestrator's chokepoint ensures every action is authorized, audited, and killable. No agent can bypass it.

Tier 1 — Sensors
Innate immunity, always on
Scout
Multi-scanner sensor (Semgrep + CodeQL + Snyk)
Oracle
PR-time risk prediction before merge
Fuzzer
Adversarial input generation on shadow-prod twin
Tier 2 — Triage
The brain
Triage
Cluster, score, deduplicate every finding
Cartographer
Live cross-repo causal graph across 5 languages
Economist
Dollar value + revenue-at-risk per bug
Tier 3 — Response
Adaptive immunity
Quarantine
Flip flags / trip breakers to contain in seconds
Blue
Writes the fix — minimal diff in isolated worktree
Red
Adversarial test writer — tries to break every fix
Judge
Ship / iterate / escalate verdict
Verifier
Runs the full gauntlet — unit, e2e, lint, SAST, perf
Prover
Formal verification (TLA+/Lean) for critical paths
Tier 4 — Memory
Immune memory
Historian
Semantic bug memory — same pattern never re-triaged
Learner
Per-class + per-signal calibration from outcomes
Tier 4.5 — Unprecedented
The 100-year moat
Dreamer
Overnight speculative refactors in parallel sandboxes
Archaeologist
Maps dormant bugs not yet triggered by any user
Negotiator
Resolves cross-team contract bugs simultaneously
Pessimist
Writes the postmortem of the outage before it ships
Time Traveler
Deterministic record-replay to prove root cause
Intent Keeper
Flags drift between what code does vs what it should
Extinction
Certifies bug classes as mathematically eliminated
Provenance
Links every line to ticket, author, and intent
Shadow Truth
Parallel implementation for differential testing
Freshness
Enforces expiration — stale code auto-quarantined
SLO Compiler
Turns plain English SLOs into hard PR gates
Tier 5 — Governance
The chokepoint
Policy
Default-deny authority matrix — veto on every agent
Auditor
Hash-chained immutable log of every action taken

Connects to your stack

BISO plugs into what you already use. No rip-and-replace.

GitHub
PR auto-commenting + merge gating via GitHub App
Semgrep
SAST scanner — findings normalized and deduplicated
CodeQL
SARIF-based analysis with cross-scanner validation
Snyk
Dependency + code scanning in native and SARIF format
Slack
BLOCK verdicts, quarantine alerts, fix loop notifications
PagerDuty
Critical pages for kill switch + quarantine events
LaunchDarkly
Quarantine flips feature flags to contain threats
PostgreSQL
Production storage with pgvector for semantic search
339
backend tests
18
live dashboard pages
57
API endpoints
12
autonomous agents
5
languages supported
3
deployment paths

Up and running in under 10 minutes

Self-hosted. Your code never leaves your infrastructure. Docker Compose brings up the entire stack — backend, dashboard, and all seven agents — in a single command.

1

Get access

Book a call or sign up — we'll set up access to the private repo and walk you through the platform.

2

Docker compose up

One command deploys the full stack to your infrastructure. No cloud dependency. Your code stays yours.

3

Click 'seed demo data'

The dashboard populates every screen instantly — cartographer graph, oracle predictions, bug inbox, metrics.

Ready to shrink the backlog?

Self-hosted and deploys in under 10 minutes. Your code never leaves your infrastructure. No cloud dependency.