Jonathan Wagner

BS"D DeepKeep, a startup among the recent winners of EIC Accelerator, develops algorithmic security and algorithmic trustworthiness solutions for AI systems. Primarily addressing LLMs and image analysis models we aim to provide: Algorithmic risk assessment of AI systems (e.g. is the model vulnerable to evasion/jailbreak attacks that force it provide illegal/harmful information? are there any trustworthiness issues ? is the model unfair? does the LLM leak sensitive private data? etc.) AI firewall solution (e.g. additional real-time layer that enhances algorithmic security and trustworthiness) including detection and mitigation of evasion/jailbreak attacks, handling data leakage etc. AI regulation compliance evaluation/enhancement These solutions handle multiple issues including: AI systems performance Evaluation Sensitive data leakage prevention Hallucination handling LLM guardrails (securing alignment of the LLM to certain topics), Prevention of malicious jailbreak/prompt injection attacks. Image data augmentation using GenAI Vision systems’ failure modes analysis and weak spot detection Evasion attacks on vision systems Fairness and Bias issues And more