Thugir Labs

Autonomous Security Research.

Proof-gated findings across web, API, network, M365, and code.

( Apply for beta ) ( How it works )

Find real vulnerabilities.
Prove them before you report them.

Authorized research across web, API, network, M365, and code — every High or Critical blocked unless exploit proof exists in code.

Built for authorized pentesters, security teams, and MSPs on their own infrastructure.

( Apply for beta ) ( How it works )
( Our process )

A research run
in 3 steps.

From scoped target to verified findings — the way a skilled pentester would, continuously.

Ready in minutes.

1

Scope the surface

Point Thugir Labs at authorized targets.

No manual checklists.
No siloed scanners.
No copy-paste between tools.

Just define what to research

Authorized scope

https://app.example.com · api.example.com · 10.0.0.0/24 · ./backend
01Web applications
02REST & GraphQL APIs
03Network services
04Microsoft 365
05Source code
06Attack chains
6 surfaces mapped Entra · repos linked
2

The research loop runs automatically

It maps the attack surface, hypothesizes, plans stages, and coordinates specialists across surfaces.

All before you review a single finding.

Planner Plan / api.example.com
Full authorized research on api.example.com — web, API, and linked code repos.
Reading openapi.json · package-lock.json
Mapping attack-surface graph · 847 nodes
Penetration task graph12 stages queued
Hypothesis: BOLA on /users/{id} — routing InjectionSpecialist + API Engine. ChainSynthesiser probing SSRF → internal metadata.
Queueing WebIntel · ServiceAnalysis · nexus_code
Reflecting · updating trajectory memory
Describe the next objective…
3

Evidence gate verifies severity

Every High or Critical goes through exploit confirmation. Unverified findings are demoted in code.

Only proof-tier findings reach you.

evidence_gate.py · precision mode
tier=01 · LLM hypothesis → demoted to informative
tier=02 · probe unconfirmed → blocked
tier=03 · tool_detected · no validation → informative only
tier=04 · exploit_verified · BOLA /users/1847 → permitted
1 critical permitted 3 demoted · precision mode
( Cognitive runtime )

Seven layers. One evidence gate.

The engine behind the work — typed event bus, Penetration Task Graph, Model Abstraction Layer. Models are workers; severity is decided in code.

( Research loop )

ObserveScanners, crawlers, and plugins register evidence
ReasonKernel interprets signals across the graph
HypothesizeAttack paths ranked by impact
PlanDynamic stages dispatched to specialists
ExperimentProbes and exploit attempts run in scope
ReflectResults critiqued against prior runs
LearnMemory and routing profiles updated
RepeatLoop continues until budget or coverage
critical_precision1.0
false_critical0
web8012/12
cve_bench_mock4/4
bountybench2/3
surfaces6
chains3 multi-surface
recall_est1.0

Specialist plane

Multi-agent reasoning over shared evidence — not a single prompt at the end of nmap.

ChainSynthesisermulti-hop attack paths
InjectionSpecialistSQLi · XSS · SSRF · traversal
WebIntelsession-aware crawl · probes
ServiceAnalysisCVE · EPSS · KEV
AD intelSMB · LDAP · WinRM
API Engine v3BOLA · JWT · GraphQL
Validationconsensus · adversarial
nexus_codeSAST · secrets · fuzz
M365 bridgeEntra · misconfig
tcsf/kernel/research_loop.py
tcsf/exploit/confirmation_loop.py
nexus/validators/evidence_gate.py
nexus/specialists/chain_synthesiser.py
tcsf/mal/routing_profiles.py
tcsf/benchmarks/mythos_bar.py
( Evidence plane )

No High or Critical without proof.

False positives waste remediation cycles. Severity is demoted in code — not in a prompt.

TierProof classHigh / Critical
01LLM hypothesisblocked — demoted
02tool probe, unconfirmedblocked in precision mode
03tool_detected, no validationinformative only
04exploit_verified + proofpermitted
05grader · sqlmap confirmedpermitted
( Current state )

Where we are today

Live

  • Research kernel on every authorized run
  • Evidence gate · impact validator
  • Web, API, M365, network, code
  • Agentic chat — 25 MCP tools
  • Golden precision 100% · 0 false criticals

In progress

  • Closed beta onboarding
  • Live CVE-Bench Docker subset
  • Weekly power cycles · Mythos Bar
  • LoRA flywheel on verified episodes

Not yet

  • Hosted multi-tenant SaaS
  • Self-serve pricing
  • Custom foundation model
  • Autonomous remediation
( Research )

Research notes on autonomous security

Written as we build — not marketing copy.

  • 2026-06-27architecture

    Severity is a code decision, not a confidence score

    How TCSF blocks High and Critical findings unless deterministic exploit proof exists — enforced in evidence_gate.py and impact_validator.py, not in a prompt.

  • 2026-06-27benchmarks

    Mythos Bar: what we measure and why

    Official scoreboard for the power stack — critical precision, Web80 exploited, CVE-Bench mock chains, recall estimate, and multi-surface chain depth.

( All research notes ) ( RSS )
( Closed beta )

Apply for access

Free access on your own infrastructure. We review every submission manually.

Or [email protected]

( FAQ )

Common questions

What does Thugir Labs do?+
Authorized penetration research across web apps, APIs, networks, Microsoft 365, and code — with exploit verification before High or Critical severity is reported.
Is this a scanner?+
No. We run a research loop. Scanners register as observation capabilities in the capability registry.
What model does it use?+
Frontier models via MAL — primarily Claude. LoRA on verified episodes. No custom foundation model.
What does beta include?+
Free closed beta. Self-hosted deployment with the full research stack. Manual onboarding.
How is this different from AI security chatbots?+
Severity is decided in code. The evidence gate blocks High and Critical unless deterministic exploit proof exists — not model confidence.