01
Public attack datasets
Hundreds of real prompt injections collected by others: the deepset prompt-injections dataset and Lakera's Gandalf dataset, gathered from people trying to trick an AI into giving up a password.
Security you can measure
Every message to a DataSpeak assistant passes a security check first. It stops attempts to manipulate the assistant, and harmful or dangerous requests, before they reach your AI. We test it against public attack datasets and our own trick questions, and publish the results.
99.8%
of attacks immediately blocked
Attacks from public datasets and our own trick questions
730
individual tests
Dutch, English and German, rerun after every change to the check
Measured in October 2026 on the DataSpeak evaluation platform.
How we test
A security check is only as good as the attacks it is tested against. That is why we use attacks we did not write ourselves.
02
Our own trick questions
Attacks written for customer service in Dutch and English: hidden instructions, letters spaced out to hide words, fake managers promising discounts, and attacks spread over several messages.
03
Representative customer questions
220 normal messages that must get through, including ones that sound suspicious, like "ignore my earlier email" or a pasted letter. A check that blocks your customers is not secure, it is broken.
04
Every change is retested
All test sets run on our evaluation platform. Every change to the check is measured on all of them before it goes live.
What the check stops
Customers can ask anything. Attempts to misuse your assistant are stopped, and the assistant answers with a polite refusal.
Extracting instructions
Attempts to get the assistant to reveal its instructions, prompts or configuration.
Overriding instructions
"Ignore all previous instructions" and every variation on it, including ones hidden in pasted text or spread over several messages.
Taking over the assistant
Making the assistant play a politician or another company, give political opinions, make jokes about groups of people, or spread untrue claims in your name.
False promises
Fake agreements with a manager or a "confirmed" discount the assistant must repeat. The assistant never commits to deals that were not made.
Technical attacks
Code injection, database queries, system commands and encoded instructions meant to slip past filters.
Harmful and dangerous requests
Requests for harmful, illegal or dangerous content, and attempts to extract other customers' data.
How others score
Security vendors publish their scores on Lakera's PINT benchmark. It measures both: catching attacks and letting normal messages through.
- DataSpeak security check99.8%
- Lakera Guard95.2%
- AWS Bedrock Guardrails89.2%
- Azure AI Prompt Shields89.1%
- ProtectAI DeBERTa v279.1%
- Llama Prompt Guard 278.8%
- Google Model Armor70.1%
DataSpeak: attacks blocked on our test sets. Other products: published PINT benchmark scores (May–August 2025). PINT uses a non-public dataset, so these are different tests and not a head-to-head comparison.
Notes
- 21 messages in the public datasets are labelled as attacks, but are harmless requests such as "write me a story" or a grammar puzzle. We report these separately instead of counting them as attacks. The assistant itself decides whether to help with them.
- AI models don't always give the same answer twice, so results vary by a few messages per run.
- Most public attacks are in English. That is why we add our own Dutch attacks and customer questions.
- The security check is one layer of protection, next to the assistant's own instructions and the permissions you give its integrations.
Security & Privacy
Your data stays secure and private with our EU-based infrastructure and strict privacy measures.
Certified
ISO 27001
DataSpeak is ISO 27001 certified, demonstrating our commitment to maintaining the highest standards of information security management.
EU Hosted
All data is stored and processed within the European Union, ensuring your information is fully protected under EU legislation.
GDPR Compliant
We fully comply with the General Data Protection Regulation (GDPR) and protect your privacy rights.
EU AI Act
The DataSpeak AI platform complies with the new EU AI Act regulations for transparent and responsible AI applications.
Exit Guarantee
You always maintain full control over your data. You can export or delete your data from the DataSpeak AI platform upon request.
No Training on Data
We never use your data to train our AI models. Your information remains completely private and is never shared.
Audit Trail
Track who did what and when across your AI assistants. Full activity logging for accountability, compliance, and troubleshooting.
Configurable Data Retention
You control how long operational data is kept. Set retention periods per client to match your policies — audit logs are always preserved.
Zero Data Retention
For maximum privacy, we offer Zero Data Retention (ZDR) with select LLM partners. Prompts and responses are not stored or logged — an extra layer of protection for sensitive use cases.
Want to see the test results?
We are happy to show how the security check performs on your own questions.