Asaf Shabtai

Failure-mode contextual bandits can boost model accuracy by over 4% on standard benchmarks while eliminating the need for additional human annotation.

Event-centric explanations can transform cybersecurity investigations by providing actionable insights that improve alert handling and classification accuracy.

Automated adversarial mitigation can now disrupt attacks while preserving network functionality, achieving a 46.7% success rate in real-world scenarios.

Forget generic chatbots – SecMate slashes cybersecurity troubleshooting failures by 40% simply by adding device-specific diagnostics.

Multi-agent code editing with structured failure feedback boosts task success by 17%, suggesting a promising path to more reliable LLM-driven code manipulation.

Gamified mobile sensors can turn cybersecurity negligence into teachable moments, boosting security awareness where it matters most: in users' everyday actions.

Code Execution MCPs slash token usage and latency in agent systems, but open a Pandora's Box of new attack vectors, demanding layered defenses.