Abhay Sheshadri

@abhayesian

AI Safety/Security Researcher

Joined August 2016

1.4K Following 624 Followers 166 Posts

Latest posts 5

New Redwood Research post

LLMs can use RL training on one task to teach themselves a different, unrelated capability. In a form of exploration hacking, the model outputs answers that get reward only when it succeeds at a side task it wants to learn. We call it reward laundering.

Abhay Sheshadri @abhayesian Mar 10

New Anthropic Fellows research: Alignment auditing—investigating AI models for unwanted behaviors—is a key challenge for safely deploying frontier models.

We're releasing AuditBench, a suite of 56 LLMs with implanted hidden behaviors to measure progress in alignment auditing.

Abhay Sheshadri @abhayesian Dec 13, 2025

🧵 Earlier this year, Anthropic ran an auditing game where teams of researchers investigated a model with a hidden objective.

Now we're releasing an open-source replication on Llama 3.3 70B as a testbed for alignment auditing research.

Abhay Sheshadri @abhayesian Jul 9, 2025

Some additional fascinating findings from our alignment faking research that didn't fit in the main thread 🧵

Anthropic Verified account @AnthropicAI Jul 8, 2025

New Anthropic research: Why do some language models fake alignment while others don't?

Last year, we found a situation where Claude 3 Opus fakes alignment.

Now, we’ve done the same analysis for 25 frontier LLMs—and the story looks more complex.

View quoted post on X

Abhay Sheshadri @abhayesian Dec 13, 2024

Excited to share this work from my internship at @CHAI_Berkeley! Grateful to have collaborated with such brilliant colleagues on this research. Thank you to my amazing co-authors! 🧵👇

Luke Bailey Verified account @LukeBailey181 Dec 13, 2024

Can interpretability help defend LLMs? We find we can reshape activations while preserving a model’s behavior. This lets us attack latent-space defenses, from SAEs and probes to Circuit Breakers. We can attack so precisely that we make a harmfulness probe output this QR code. 🧵

View quoted post on X