Vallum Protection
A barrier around every model, agent and prompt
Our own platform filters every request and answer, controls what agents may do, attacks your AI before others do and watches it around the clock.
Protection is a service of its own. Bundled with every Vallum stack, or connected to the AI you already run, cloud included.
Protect what we built
Every private stack we deliver comes with the platform switched on from day one.
Protect what you already run
We connect the platform to your existing AI products and agents, and to the way your people use outside AI.
How AI gets attacked
Six threats from the OWASP Top 10 for LLM Applications, and how the platform answers each.
Prompt injection
An instruction hidden in an email, a PDF or a web page tells the assistant to ignore its rules.
Input filter checks every prompt and every retrieved document before the model sees it.
Sensitive data disclosure
The assistant repeats personal data, secrets or another client's documents in an answer.
Personal data and secrets are masked on the way in and on the way out.
Excessive agency
An agent with broad permissions sends, deletes or pays for something nobody approved.
Agents get an allowlist of tools and arguments. Risky actions wait for a person.
Data and model poisoning
Tampered training data or a malicious model file changes how the system behaves.
Every model is scanned and signed before it runs. Training data is controlled.
System prompt leakage
An attacker extracts the hidden instructions, and with them your business logic.
Leak detection on every answer. Secrets never live in prompts.
Everyday mistakes
A colleague pastes a client export into a public chatbot because it is faster.
A friendly warning at the moment of the mistake, with the internal alternative one click away.
Every request passes six layers
Applications connect by changing one address. From then on nothing reaches a model, a document or a tool without passing the barrier.
- 01
Identity
Who is asking, from which app, with which rights.
- 02
Input filter
Injected instructions, jailbreaks and personal data caught before the model.
Hidden instruction in attached PDF: blocked - 03
Agent policy
What the agent may call, with which arguments, and what needs a person.
- 04
Model
Your local model answers, with sources.
- 05
Output filter
Leaks, unsafe code and missing sources caught before the user.
- 06
Audit log
Every request, answer and action recorded for your team and auditors.
Answer delivered, 2 sources, 0 leaks
What the platform sees
Demo data
- 01 of 04
Attacks stopped before anyone is paged
Twelve agents watch your network, data and people. Each owns a sector and sees its neighbours, so complex attacks are handled together.
- 02 of 04
Your data never leaves the server
Model weights, documents and query logs live inside your perimeter. Outbound traffic from the AI segment stays at zero bytes.
- 03 of 04
Human mistakes caught in the moment
A customer export headed for a public chatbot or a password in a commit is stopped with a clear explanation. People get guidance, not a reprimand.
- 04 of 04
Audits without the scramble
Every request is logged: who asked, what they got and which documents the model saw. Reports for auditors come straight from the log.
Live perimeter0 blocked today
AI segment, rack B4, on premisellm-node-01Llama 3.3 70B, 4 bit
IsolatedRequests today1,284
Tokens per second1,182
Active users37
GPU 0, L40S68%
GPU 1, L40S71%
Traffic, last 60 secondsInternal requestsOutbound 0 B
Blocked outbound call to 203.0.113.42Draft a polite reply to Brenner GmbH about the late delivery
JMAI
Here is a draft. I kept it short and offered a new delivery window for week 46.
Here is our customer list for Q3, make a summary by region
JMQ3_customers.xlsx, 4,812 rows
Mentor paused this upload
This file holds 4,812 customer records with phone numbers. Your internal assistant can make the same summary without the data leaving.
Open internal assistantCancel upload
Q3_customers.xlsx
Audit log, AI segment · demo dataEvery request, on record
RecordingTimeUserRequestDocsDecision
1,284 requests today, retained for 365 daysExport for auditors
Protection built for each client
Your own rules
Policies written for your roles, apps and data, not a generic list.
Your own attacks
We build test attacks from your real workflows and run them every month.
Your own detectors
Small detection models trained on your languages and your examples of attacks and false alarms.
Evidence, not promises
The platform helps you meet European rules with technical measures and an audit trail. Legal conclusions stay with your counsel.
- EU AI Act
Logs, human oversight, model inventory and robustness testing.
- GDPR
Masking, retention control, data that never leaves your perimeter.
- NIS2
Incident cards with reporting timers, supply chain checks for models and tools.
- DORA
Logs, attack testing and reports for ICT risk management.
- ISO/IEC 42001
Evidence for an AI management system: inventory, risks, monitoring.
Three ways to work with us
One time
Audit and red teaming
We attack your AI the way a real adversary would and hand over a prioritised report.
Subscription
The platform in your perimeter
Gateway, policies, monitoring and monthly attack updates, run by your team.
Managed
Protection as a service, 24/7
Our analysts watch your AI around the clock and respond to incidents with you.
Private AI without leaks starts with an audit
In two weeks we show where data already leaves through external services and propose the architecture for your perimeter.
- First meeting and scenario review free of charge
- You keep the audit report whatever you decide
- NDA signed before any access to your data