Hades

Adversarial AI for Security Engineers

Autonomous red-teaming,enforced at every commit.

Frontier AI has already written offensive tools that find zero-days at scale — locked behind 40-org access lists. Hades gives you the same class of defense, running inside your own cloud.

Your engineers ship fast — and so do the models writing half the code. Hades keeps pace by scanning every commit for exploitable weaknesses, from classic injection and auth bypass to prompt injection and unsafe tool chains — then proves each one with a working payload before it hits main.

Trusted by security teams at

Stripe
IBM
Samsung
Vercel
Bosch
Discord
Airbnb
Uber
DigitalOcean
Canon
Hisense
Stripe
IBM
Samsung
Vercel
Bosch
Discord
Airbnb
Uber
DigitalOcean
Canon
Hisense
hades — run #a7f3

>target acquired: src/auth/session.ts

>mapping trust boundaries...

>hypothesis: session token survives logout

>spawning sandbox [hades-a7f3]

>payload: replay stale token (T+47s)

>executing...

|SUCCESS — token valid post-logout

!impact: account takeover

!severity: critical (9.1)

>reproduction: attached

>sandbox destroyed

|

Pixel grid
Attribute
Coverage model
False positive rate
Exploit proof
Cadence
AI-era threats
SAST
Static analysis
Pattern / taint rules
High — manual triage
None — findings only
Every commit
Misses agentic flows
DAST
Black-box scanner
Surface crawling
Medium
Weak — generic payloads
Nightly / weekly
Misses prompt-level
Pentest
Human engagement
Scoped by agreement
Low
Strong — bespoke
Annual / quarterly
Expensive, rare
Hades
Autonomous red team
Adversarial reasoning
Low — proof-gated
Working payload, every finding
Every commit
Native: prompt injection, unsafe tool chains, agent abuse

Hades is an autonomous red team. It reasons adversarially about your codebase the way a skilled attacker would — enumerating trust boundaries, chaining weaknesses across services, and synthesising working exploit payloads against live targets inside your own environment. The output is not a list of suspects; it's a set of reproducible zero-day-class findings with the exploit path, the payload, and the patch that closes it.