Our product

Our product protects the automated processes from manipulation or misuse by an attacker. The functionality is divided into three complementary modules.

The Input Security module provides real-time checking and sanitization of inputs to let the customer’s model concentrate on its business goals.

The Model Oversight module continuously monitors the decisions made by the protected model to discover manipulation, strategic behaviour or systematic false positives.

Active Model Assessment identifies the vulnerabilities and possible bias in the protected system.

Group 3 Created with Sketch.

Our Customers

Finance & Fintech

Finance and Fintech companies have been using AI, machine learning, statistical decision making and expert systems in order to provide secure and convenient services to their customers. The protection of automated decision-making is especially vital in scenarios lacking humans in the loop. Bulletproof AI supports its financial sector customers in protecting their existing:

  • credit risk management and instant credit systems
  • fraud detection & prevention
  • anti-money laundering systems
  • adaptive and rule-based authentication & authorization

Bulletproof AI monitors and constantly probes the behavior of the protected system, recognizes potentially evasive, artificial or manipulative behavior and provides testing and training data necessary for system hardening against the detected attacks.

Our customers can use Bulletproof AI to prevent novel fraud scenarios that may arise due to business changes associated with PSD2 regulation and the introduction of 3D Secure 2 and Strong Customer Authentication standards.

Input Checking

Input Security module prevents the suspicious inputs from entering the protected system. It directly prevents losses by blocking malicious transactions and makes the attacker's life more difficult as it limits the information leakage from the protected system to the attacker.

In-Line Transaction

Input Security Module can be tightly coupled with the Model Oversight Module. It can use the long-term, cross-transaction scores from the Oversight module to block the new transaction associated with past problematic behavior.

PDF Forgery

Detection of forged and manipulated PDF documents protects automated processes that rely on unauthenticated documents received from third parties. Consumer and small business lending, consumer and car financing and factoring solutions are the typical users.

Model Oversight

Model oversight module identifies the advanced attacks, strategic model probing and manipulation. We protect the underlying model by analysing the whole stream of transaction queries instead of the individual transactions. This gives us the additional context necessary both to detect the additional attacks, as well as to de-emphasise probable false alerts created by the underlying system.

False Positive

False alerts and unjustifiable denials are drain on customer's business. The manual work necessary for their effective inspection is expensive and analysts frequently complain about the low relevance of the alerts they receive. Our system prioritises the alerts from the underlying system and connects them together to drive actionable outcomes.


Automated decisions are a factor in the life of today's attackers. Knowingly or just by intuition, the attackers can identify the vulnerabilities in the automated decisions and exploit them at scale. Model oversight module discovers early-stage symptoms of the attacks provides an early warning before a substantial loss occurs.

Active Model Assessment

Active Model Assessment continuously probes the protected system with attack-like inputs based on module's knowledge of the protected model and the input data. It verifies whether the outputs match the expectations in order to assess the security and fairness of the system.


Bulletproof AI can be used in a one-shot mode to deliver an instantaneous security assessment of the system in gray-box and black-box settings. The outcome of testing can be used for system re-training, for machine learning model hardening or for robust feature engineering.


Fairness assessment analyzes the model and discovers any instances, where the lack of data or bias in feature selection, may have biased the model response. The assessment includes an analysis of the model's past performance extended by hypothetical scenarios generated by Bulletproof AI.

Model Pre-Deployment

The Bulletproof AI solution can easily be deployed as part of the model release pipeline. When a new model is released, it can be assessed from a security and fairness perspective in order to ensure proper behavior.

Our team

The Bulletproof AI s.r.o. team is based in Prague and Brussels. We have created Bulletproof AI with 12 years of research and industry experience in security and machine learning. Research team members have authored or co-authored hundreds of patents, journal articles and conference papers in the field of machine learning and security. Our previous startup, Cognitive Security, was acquired by Cisco Systems in 2013 and has grown to protect more than 25 million of users worldwide from advanced malware threats using machine learning.

Martin Rehak


Jan de Waard

Business Development

Jan Doubek

Customer Solutions

Karel Bartoš


Martin Grill


Jan Jusko


Tomas Laube

ML Engineer

Lukas Machlica


Josef Stach

Security Engineer

Jan Stiborek


Martin Vejman

ML Engineer


Are you interested?