AIM Intelligence

On July 7–8, during ICML 2026 week, AIM Intelligence co-organized the frontier safety and red-teaming program at the Seoul Forum on AI Safety & Security 2026 (SFASS). Hosted by Korea’s Ministry of Science and ICT and organized by ETRI and the Korea AI Safety Institute , the forum was held in collaboration with MLCommons, the Future of Life Institute (FLI), AIM Intelligence, Microsoft, and Google. Across two days, leaders from government, research, and industry explored the next frontier of AI safety and security, from global standards and multilingual evaluation to multimodal models, agents, and physical AI. AIM contributed across the forum’s emerging-risk and red-teaming program. Our Co-founder and CTO, Haon Park, presented: “Frontier AI Safety and Security: Agent, Multimodal, and Physical AI” AIM also led the Red-Teaming Workshop, bringing together experts from Google, Google DeepMind, Microsoft, Korea AI Safety Institute, FAR.AI, LG AI Research, Singapore AI Safety Institute, ELLIS Institute Tübingen. The forum concluded with AIM’s live challenge: Judgement Day: Jailbreaking the Humanoid Participants attempted to manipulate an AI-controlled humanoid into performing unintended physical actions, demonstrating a critical shift in the threat landscape: A jailbreak is no longer only a harmful response. It can become a harmful action. The future of AI safety must protect the entire system, from perception and reasoning to tools, permissions, runtime control, and physical execution. Thank you to every organizer, speaker, participant, and AIM team member who made SFASS 2026 possible.

New paper: EgoSafetyBench, a diagnostic benchmark for evaluating VLMs as runtime safety guards for embodied agents. Vision-language models are increasingly being proposed as the safety layer for robots operating in homes and factories. But there's a hard question underneath that: can these models actually tell a genuinely dangerous moment from one that just looks alarming? A kitchen knife isn't inherently unsafe. It depends entirely on what the robot is doing and the context around it. Most embodied-safety benchmarks flatten all of this into a binary safe/unsafe label, which hides exactly the failures that matter in deployment. So we built EgoSafetyBench: 1,200 egocentric robot-view video scenarios, annotated at half-second granularity, where VLMs are evaluated as streaming guards judging each moment as it happens. One track covers situations ranging from routine activity to safe-but-suspicious scenes to obvious and contextual hazards. The other targets in-scene text, like a sticker or sign that lies about what's actually happening. Every misleading sign has a truthful control on an identical scene, so any change in the model's behavior comes from the deception itself. We tested 10 open and closed-source VLMs. Three things stood out: Models are good at noticing that a video contains a hazard somewhere. They are much worse at catching the exact hazardous moment, especially when the danger depends on context. Misleading text degrades every single model, but in opposite directions. It talks vulnerable open models out of real hazards, while it scares the strongest closed models into flagging scenes that are perfectly safe. A lot of what looks like robust safety in top models is actually indiscriminate alarming, not real physical reasoning. Without the matched controls, we would never have caught this. The bigger point: if we want deployable safety guards, we need to evaluate false-positive control, contextual reasoning, and resistance to deceptive channels, not just aggregate accuracy. Great working on this with Siddhant Panpatil , Mijin Koo, Chaeyun Kim, Haon Park, and Dasol Choi at AIM Intelligence and Seoul National University accordingly. Paper: https://lnkd.in/ghHBqAEW Code and data on GitHub and HuggingFace.

AIM Intelligence is heading to LEAP East 2026 in Hong Kong. As AI moves beyond chat into agents, multimodal models, and physical AI, enterprises need stronger ways to evaluate, monitor, and secure AI systems before real-world deployment. At AIM Intelligence, we help organizations deploy AI safely through AI red-teaming, guardrails, risk/compliance assessment, and continuous monitoring across frontier models, agents, and enterprise AI applications. We are excited to meet global partners, customers, and research collaborators at LEAP EAST 2026. 📍 Visit us at Booth H3.O78 You can also find AIM Intelligence on the LEAP EAST app and contact us directly to schedule a meeting. Attending from AIM Intelligence: Haneul Kim (Skye) — CFO, Co-Founder Hanwool Lee — Head of Defensive Security 강태웅 KANG  — Technical Staff of Offensive Security See you in Hong Kong! #LEAPEAST2026 #AIMIntelligence #AISafety #AIRedTeaming #AgentSecurity #HongKong

Heading to Seoul this July for ICML? Make sure to check out the Seoul Forum on AI Safety & Security (SFASS)! Hosted by Korea AI Safety Institute, this event brings together global leaders and researchers to advance AI safety. We at AIM Intelligence are participating as a collaborating partner, bringing you hands-on programs to test and secure the future of AI. 𝐑𝐞𝐝 𝐓𝐞𝐚𝐦𝐢𝐧𝐠 𝐖𝐨𝐫𝐤𝐬𝐡𝐨𝐩 - This workshop dives deep into the latest in AI Red Teaming, covering practical methodologies for probing LLM vulnerabilities, multi-modal foundation models, and engineering safety guardrails. - Featuring an incredible lineup of speakers and panelists: Jenny Ni ✨, Naman Goyal, Eugenia Kim, Sang Seo, Dasol Choi, Maksym Andriushchenko, Haon Park, Myoung-Shin Kim, and Adam Gleave.  𝐋𝐢𝐯𝐞 𝐉𝐮𝐝𝐠𝐞𝐦𝐞𝐧𝐭 𝐃𝐚𝐲 𝐂𝐡𝐚𝐥𝐥𝐞𝐧𝐠𝐞 - Date: July 8th (1.5 Hours) - Following our recent online challenge, we are bringing the competition to the physical stage. Watch top minds compete in a real-time AI safety hackathon. In partnership with MLCommons, Future of Life Institute (FLI), Microsoft, Google See you in Seoul! https://lnkd.in/gT-a6w7y

🚀 AIM Intelligence is joining Microsoft's MDASH limited preview as an Early Partner. We are honored to be the only publicly listed Korean company participating alongside 48 global leaders including IBM, Accenture, PwC, Infosys, and TCS in shaping the next generation of AI security. As AI rapidly evolves from chatbots to autonomous agents and Physical AI, security must evolve as well. MDASH represents Microsoft's vision for agentic AI security: leveraging multiple specialized AI agents to autonomously discover and validate vulnerabilities at scale. It recently topped the public CyberGym benchmark, surpassing the published preview result of Anthropic’s Mythos. We are excited to contribute our expertise in AI Red Teaming, Agent Security, and Runtime Guardrails to this ecosystem. At AIM Intelligence, we believe the future of AI security goes beyond protecting models. It is about securing AI systems that can perceive, reason, and act autonomously in the real world. This collaboration further strengthens our mission: Keep AI Under Control. Microsoft MDASH Engaged Partners: https://lnkd.in/gfQuaSNS

Meet AIM Intelligence at VivaTech 2026 in Paris. From AI applications and autonomous agents to physical AI, organizations need more than visibility: they need the ability to test, secure, and control how AI behaves. At VivaTech, we’ll showcase AIM Intelligence’s technical foundation for AI safety and security: 🔴 Stinger: automated red teaming for AI models, agents, multimodal systems, and physical AI 🔵 Starfort: real-time guardrails and runtime protection for enterprise AI ⚙️ Physical AI Guard: an interpretable control layer designed to intervene before unsafe physical actions occur Meet our team in Paris: Haon Park | CTO & Co-Founder Chaewook Kim | GTM Principal Manager Mijin Koo | AI Researcher 📅 June 17–20, 2026 📍 Paris Expo Porte de Versailles 🎪 June 17–18: Orange Booth, Pavilion 7 - Hall 7.2, 2D16-001 Schedule a 15-minute meeting: https://lnkd.in/gwJ_jKiU Let’s discuss how to keep AI under your control.