Arth Singh

Arth Singh

AI Safety Research Scientist, AIM Intelligence

Research Assistant, Jinesis Lab (University of Toronto and Vector Institute), with Zhijing Jin

I study how frontier and agentic models behave when something is off: reward hacking, emergent misalignment, sycophancy, and monitor evasion. Most of my work builds evaluations that surface these behaviors and then tries to explain them, from the attack side and from inside the model.

My other major interest is AI informed by neuroscience: building systems that learn from how the human brain works and behaves. I am also working on systems inspired by biological organisms, to see which of their mechanisms can carry over to AI.

I am an AI Safety Research Scientist at AIM Intelligence in Seoul, where I set the frontier safety research agenda after working on engineering for the AI red team, and a research assistant in Zhijing Jin’s Jinesis Lab at the University of Toronto and the Vector Institute. I am studying for an online M.S. in Computer Science (Machine Learning) at Georgia Tech.

I studied biotechnology and biochemical engineering at NIT Agartala and moved to AI safety in my third year, coming in through red-teaming: AI bug bounties on Mozilla’s Ødin and pre-release testing of Anthropic’s Constitutional Classifiers.

I was born on 2 November 2004 in Mumbai, India, and now live in Seoul, South Korea. I am 21 and an INFJ.

Experience

  • AIM Intelligence, Seoul

    • AI Safety Research ScientistSep 2026 – present
    • AI Safety Research EngineerOct 2025 – Aug 2026

    Sets AIM's frontier safety research agenda: evaluations that surface and explain reward hacking, emergent misalignment, sycophancy, and monitor evasion in frontier and agentic systems. Previously led red-teaming engagements for clients including LG, Samsung Group, and Shinsegae, expanded the attack coverage and evaluation pipeline of AIM's automated red-teaming platform (Stinger), and worked on its guardrails (Starfort), including audio guardrails.

I am a two-time SPAR fellow. In Fall 2025 I worked with Kellin Pelrine and Tom Tseng at FAR.AI, building parts of their automated red-teaming pipeline for biological-harm evaluations. In Spring 2026 I worked with Rishub Jain of Sampura Research on human-AI complementarity for identifying harmful conversations.

I was also a MARS 5.0 fellow at the Cambridge AI Safety Hub, working with Samuel Simko and Zhijing Jin on causality-based adversarial defenses for open-weight models.

I have also worked with Joshua Jacob of Sampura Research, and with David Africa and Konstantinos (Kozzy) Voudouris of the UK AI Security Institute.

Highlights

  • I started doing research in February 2026 and have 10 papers since, 6 of them accepted at venues including NeurIPS 2026 and Findings of EMNLP 2026.
  • Part of the AIM Intelligence team that bypassed the safety guardrails of Google’s Gemini 3 Pro in five minutes (Maeil Business Newspaper, November 2025).
  • Letter of recognition from NASA’s Vulnerability Disclosure Program for reporting a vulnerability in a NASA-affiliated LLM endpoint tied to the Parker Solar Probe. The endpoint was taken offline.
  • Reached #1 globally on the Mozilla Ødin AI bug-bounty leaderboard, earning more than $25,000 in bounties.
  • Only contributor from the APAC region to OpenAI’s agent guardrail red team; the findings were presented at OpenAI DevDay.
  • Contributed 30 jailbreak templates to PyRIT, Microsoft’s open-source red-teaming toolkit (merged December 2025).

Recent writing

All posts

Selected publications

All publications and preprints

Service

Reviewer for AI4GOOD 2026 (NeurIPS), TAIGR Workshop 2026 (ICML), AI4GOOD 2026 (ICML), and TrustNLP 2026 (ACL).