ERA:AI Fellowship Summer 2026

At AI Safety Hungary I learned that RLHF - built to align models - became the primary driver behind GPT-4's capabilities. Safety research was feeding the race it was meant to slow. Since then I: - Founded AI Safety India, ran Technical AI Governance cohort training 30 people - Top 30 at Grayswan cracking 13 models using multi-turn encoding attacks - SPAR project building a Capability Spillover Potential framework - early analysis - interpretability scores higher on spillover risk than most researchers assume - 450 experiments on inter-agent safety degradation across three Gemini generations. Finding: subtle helpfulness-optimized framing produces twice the behavioral variance of an explicit jailbreak persona. Submitted to ACL 2026 SRW. The problem I want to solve in ERA: India's ONDC and UPI handle 18 billion transactions/monthly and ₹49 trillion in welfare routing with no evaluation methodology for failure modes. When these systems fail, welfare exclusion is systematic and unaccountable at a scale no other country produces. Kirch and Field's ERA 2024 work is the closest foundation - but it stops at single-model surfaces. Nobody has built the multi-agent version. That's what I want to work on.

Recently, I have run 450 experiments testing whether safety-aligned agents degrade when exposed to misaligned ones through natural language. The finding - subtle helpfulness-optimized framing is twice as dangerous as an explicit jailbreak persona - is submitted to ACL 2026 SRW. Currently I am working on SPAR project - Capability Spillover Potential framework - tracking which safety research directions inadvertently feed capabilities. Early analysis suggests interpretability research carries higher spillover risk than the community currently prices in. I placed Top 30 at Grayswan cracking 13 models using multi-turn encoding attacks. I founded AISIN and facilitated technical AI safety cohorts through AI Safety Collab. Paper: drive.google.com/file/d/19MrEIUPFR3CCexXG0Opic0cmXmUg90bb SPAR write-up: adityarajai.substack.com/p/spar-research-capabilities-spillover

Time Horizon can measure something that MMLU cannot - End to end autonomous work a model can do without any human intervention. Core weakness is - 1. It's too narrow (Only covers software, ML and Cybersecurity) and time horizons on governance tasks and social manipulation, etc. where real deployment harm occurs are unmeasured and likely different. 2. The trend assumes compute growth continues. METR's November 2025 compute slowdown paper shows that if compute growth slows, the 1 month reliability milestone shifts from 2029 to 2033. A more informative alternative is - Tracking the rate at which AI accelerates AI R&D itself. RAND's 2026 AGI forecasting report identifies this as the leading indicator for discontinuous capability jumps - precisely the scenario time horizon extrapolation misses.

The research question I want to work on: how do you evaluate safety at the system level in multi-agent deployments, when every individual agent passes its safety checks but the collective produces harmful outcomes nobody designed? India is the right place to study this. ONDC puts a buyer's agent, seller's agent, platform's agent, and regulatory agent on a single transaction simultaneously. UPI routes 18 billion monthly transactions through agent-mediated fraud detection. DBT routes ₹49 trillion in welfare transfers. These are not hypothetical systems - they are live, at scale, and have zero evaluation methodology for system-level failures. My ACL paper showed empirically that safety failures emerge from interaction structure, not individual agent behavior. That is the empirical baseline. What is missing is the evaluation framework that catches it before people get excluded from welfare they are entitled to. Three sub-questions: What conditions produce emergent unsafe outcomes between individually safe agents? Can we build an eval methodology that catches system-level failures individual evals miss? Which governance mechanisms are tractable for MeitY given India's institutional structure? Deliverable: a methodology paper and a policy brief for MeitY. Theory of change: a replicable eval framework here becomes the template for every government running agent-mediated public infrastructure - which will be most governments within a decade.

Through a WhatsApp group for AI safety practitioners in India. After reading ERA's past research - particularly Kirch and Field's jailbreak mechanism work - I decided to apply.

Link of Research Paper and Early results on Capability spillover - drive.google.com/drive/folders/1fO_RlrSgnq_MskmcAUwRlzWIAN4scxcX?usp=sharing