Jailbreak
Role-play, obfuscation, and multi-turn escalation attempts to bypass your system prompt.
Independent AI assurance
We break your AI.
We break your AI.
Then we prove it's still safe.
Probative continuously red-teams your production AI — jailbreaks, prompt injection, PII exfiltration, tool misuse, bias, hallucination — and grades every result with an independent judge. Governance tools document your AI. We break it, and hand you the signed evidence.
The battery
Eleven ways your AI fails. Tested on repeat.
PRB-01
PRB-02
Prompt injection
Hostile instructions smuggled through user content, documents, and tool outputs.
PRB-03
PII exfiltration
Attempts to extract personal data your system holds or has seen.
PRB-04
Bias
Differential treatment across protected characteristics in identical scenarios.
PRB-05
Hallucination
Confident fabrication under pressure — citations, numbers, people, precedent.
PRB-06
Toxicity
Harmful output under provocation, adversarial phrasing, and persona pressure.
PRB-07
Refusal drift
Whether your guardrails are the same today as the day you shipped them.
PRB-08
Data governance
Leakage of system prompts, internal tools, and configuration through the front door.
PRB-09
Tool misuse
Coaxing your agent into calling the wrong tool, with the wrong arguments, on someone else's behalf.
PRB-10
RAG exfiltration
Pulling retrieval-indexed documents out through the answers they ground.
PRB-11 · agentic
Excessive agency
Whether your agent stays inside its mandate when an attacker widens it — the failure class regulators are reading about right now.
Enforcement
What enforces when.
The 2026 Digital Omnibus moved the EU AI Act's high-risk deadlines. Anyone still selling you an "August 2026 high-risk deadline" is selling a date that no longer exists. Evidence that doesn't lie starts with a timeline that doesn't either.
Aug 2, 2025
GPAI rules applied
Obligations for general-purpose AI model providers took effect.
Aug 2, 2026enforcing
GPAI fines · Article 50 · Article 4
The AI Office's enforcement powers activate: fines up to 3% of global turnover or €15M for GPAI providers. Article 50 transparency — chatbot disclosure, deepfake labeling — and Article 4 AI literacy enforce. Only Article 50(2) machine-readable marking of synthetic content slipped, to Dec 2, 2026.
Dec 2, 2027
Annex III high-risk obligations
Stand-alone high-risk systems (Arts. 9–15) — hiring, credit, education, essential services, biometrics — postponed from 2026 by the Digital Omnibus. Sixteen months of behavioral evidence is what a defensible conformity file is made of.
Aug 2, 2028
Annex I embedded high-risk
High-risk obligations for AI embedded in regulated products — machinery, medical devices, vehicles.
The vault
Evidence your auditor can verify without trusting us.
Every probe result is hashed into a per-system chain and signed. Change one record and the chain breaks — visibly, mathematically, for anyone who checks.
scan 0141sha256:9f2c…a41e✓ ed25519 signed
scan 0142sha256:1b7d…03f9 ← 9f2c…a41e✓ ed25519 signed
scan 0143sha256:c58a…7b12 ← 1b7d…03f9✓ ed25519 signed
Independent by design
Probative has no stake in your model, your cloud, or your security vendor. A judge that grades its own remediation isn't a judge.
Untested means untested
When no probe behaviorally tests a control, Probative reports it as untested — the honest coverage gap — instead of mapping it anyway.
Framework-mapped
Findings roll up to the EU AI Act, NIST AI RMF, ISO 42001, and the Colorado AI Act. One scan, four evidence packs.
The difference
Documented is not tested.
Governance platforms inventory your AI and collect attestations. Eval libraries test what your own team thought to test. Neither hands your auditor evidence they can verify without trusting you.
| Capability | GRC / governance platforms | DIY eval libraries | Probative |
|---|---|---|---|
| Adversarial behavioral testing | Questionnaires & policies | If you build it | 11 attack classes, every scan |
| Independent judge | Self-attested | Grades its own homework | No stake in your stack |
| Tamper-evident evidence | Editable records | Logs & spreadsheets | sha256 chain, ed25519 signed |
| Framework mapping | Control checklists | Manual | EU AI Act · NIST · ISO 42001 · Colorado |
| Continuous re-verification | Annual review cycle | When someone remembers | Scheduled, signed, on repeat |
Pricing
Start free. Prove it monthly.
Spot Check
$0
one scan, on us
- Full 11-class probe battery
- Independent judge grading
- Snapshot evidence report
Pulse
$450
per month · annual by invoice
- Continuous scheduled scans
- Signed evidence chain
- All four framework mappings
- Scheduled re-verification
Assurance
$1,900
per month · annual by invoice
- Everything in Pulse
- Drift alerts between scans
- Audit-grade evidence packs
- Multi-system portfolio view
- Regulated-buyer questionnaire support
Step zero
Before you prove the system behaves, know which rules apply.
Probative produces evidence against obligations you already know you have. If you don't yet know which of the EU AI Act, Colorado, Texas, Illinois, NYC, or California regimes reach your systems, start with Obligent — our sibling instrument: a deterministic, cited applicability engine that answers which obligations, from which rule, as of which date, and says needs analysis instead of guessing. Diagnosis there; evidence here.
What Probative is not. Probative is behavioral evidence, not legal advice, and not a certification body. Evidence packs show what your AI did under attack, mapped to framework controls — your counsel and your auditor decide what it means. Probative is currently in private pilot; every claim on this page describes the shipped probe battery and evidence engine, not aspirations.
GPAI fines land August 2.
December 2027 is closer than it looks.
Sixteen months of signed behavioral evidence is what a defensible conformity file is made of. Start the chain now.