Lakera, a leading AI-native security platform and a Check Point company, has launched the backbone breaker benchmark (b3), an open-source security evaluation tool designed to assess the robustness of large language models (LLMs) within AI agents. Developed in collaboration with researchers from the UK AI Security Institute (AISI), b3 introduces a novel approach called "threat snapshots" to identify vulnerabilities at critical junctures in AI agent workflows. Unlike traditional benchmarks that simulate entire agent interactions, threat snapshots isolate high-risk moments—such as when an LLM processes sensitive inputs or executes tool calls—where security flaws are most likely to emerge. This targeted method enables developers and model providers to evaluate model resilience against realistic adversarial attacks without the complexity of full-agent simulations. The benchmark incorporates 10 representative threat snapshots and leverages a dataset of 19,433 adversarial examples gathered through Gandalf: Agent Breaker, a gamified red teaming platform. This platform challenges users to exploit AI agents in realistic scenarios, simulating behaviors seen in real-world applications such as chat interfaces, code execution, file handling, memory use, and external tool integration. The game, originally developed during an internal hackathon at Lakera, has grown into the world’s largest red teaming community, generating over 80 million data points since its 2023 launch. Initial testing of 31 popular LLMs using b3 revealed widespread vulnerabilities to attacks including system prompt exfiltration, phishing link insertion, malicious code injection, denial-of-service attempts, and unauthorized tool usage. These findings underscore the urgent need for stronger security measures in AI agent development. The b3 benchmark is now available under an open-source license at https://arxiv.org/abs/2510.22620. Lakera’s team, including Co-Founder and Chief Scientist Mateo Rojas-Carulla, emphasized that the tool is designed to help developers and model providers measure and improve their security posture in a practical, scalable way. Lakera, founded in 2021 by David Haber, Dr. Mateo Rojas-Carulla, and Dr. Matthias Kraft, was acquired by Check Point Software Technologies Ltd. (NASDAQ: CHKP) in 2025. The company, dual-headquartered in Zurich and San Francisco, specializes in AI-first security for agentic applications, protecting major enterprises from emerging AI-driven cyber threats. Its security solutions are powered by real-time insights from Gandalf and proprietary AI technologies. Check Point, a global leader in cybersecurity, continues to expand its AI security capabilities through Lakera and the establishment of its Global Center of Excellence for AI Security. The company’s Infinity Platform and open ecosystem aim to deliver prevention-first protection across on-premises, cloud, and workspace environments. For more information, visit Lakera.ai, explore Gandalf: Agent Breaker, and follow the company on LinkedIn.