A adversaria Book an evaluation

ADVERSARIA · SECURITY RESEARCH UNIT OF FINIUM.TECH

We attack AI agents.
You get proof.

Adversarial evaluation for AI agents and models — prompt injection, tool abuse, data exfiltration, auth bypass. Every engagement ends in a tamper-evident evidence pack, not a slide deck.

Book an evaluation Read the method

Authorized targets only · scope-bound · human-supervised · no data retention

PROOF LEDGER

Nothing to publish yet. That is the point.

An AI-assisted hunter works on scope-bound bug-bounty and VDP programs under continuous human supervision. We publish the ledger only when facts clear two gates: program permission and technical disclosure approval. Until then: process, not promises.

  • STATUS — RUNNING
  • PUBLIC PROGRAMS MONITORED — 30+
  • FINDINGS PUBLICLY DISCLOSABLE — 0
  • PUBLICATION GATE — HUMAN + PROGRAM OWNER

No targets, counts, timestamps or payloads are published. Unknown is not zero; withheld is not failure.

What you can buy

01 · DONE-FOR-YOU

Agent Red Team

Book this →

A bounded adversarial campaign against one AI agent: prompt injection, tool-use abuse, data exfiltration, auth bypass, agent-to-agent attacks. Deliverable: findings report with reproducible PoCs and a hash-chained evidence pack your buyers can verify offline.

€4,900 fixed · per agent · 2 weeks

02 · COMPLIANCE

EU AI Act Art. 15 Evidence Pack

Robustness and cybersecurity documentation for high-risk AI systems, mapped to Article 15 requirements. Built for EU vendors preparing conformity — in English or German.

€1,900 add-on · €3,500 standalone

03 · CONTINUOUS

Continuous Assurance

Monthly adversarial regression runs against your updated agent, with delta evidence packs. Waitlist at launch; onboarded in order of request.

€290/mo · per agent · waitlist

Free, forever: the method, the open-source stack below, and the ledger.

Bounded by design

01

Scope

We test only what a signed scope or program policy explicitly allows. Allowlist, never denylist. Unsure = out of scope.

02

Attack

Passive recon and adversarial campaigns with strict rate limits and identification headers. Destructive testing never runs unattended.

03

Verify

A human reproduces every finding. Nothing enters a report on model output alone. False positives are published as ruled out.

04

Evidence

Findings ship as a tamper-evident pack: hash-chained, verifiable offline with our open-source verifier.

What we never do

No unauthorized testing · no real-user data in PoCs · no public disclosure without program approval · no fabricated telemetry on this site.

The stack is public

The rails we sell ride on are open. Audit them.

LINEAGE

Built inside an evidence company.

Adversaria is the security unit of finium.tech — the applied intelligence lab behind Finium Legal, an AI evidence desk for EU legal and comms teams. We spent years making evidence tamper-evident for lawyers; now we produce the same grade of evidence about AI systems themselves.