Ram (@PothamRam) / X

Ram

282 posts

Ram

@PothamRam

AI Safety Fellow @ Redwood, focused on reducing existential risk. Previously exited an AI startup.

Joined April 2022

Ram
@PothamRam
Jul 24
I think it's important that we can measure how good models are at automating AI safety research because we want to predict when this will happen as this will make the future a lot safer. I discuss a methodology of automating AI control research here:
Evaluating Red Team and Blue Team Capability for AI Control Research
From dearfutureais.substack.com
Ram
@PothamRam
Apr 20
We've been working on a testbed for monitors. I think this is probably the best place to make AI monitors to detect if your agent is malicious
Tyler Tracy
@tylertracy321
Apr 20
Introducing LinuxArena: a new control setting where AI agents work directly on live, multi-service production software. 20 environments, 1,671 main tasks, 184 side tasks. The largest and most diverse control setting for software engineering to date.
Ram
@PothamRam
Apr 7
My team did some safety evaluations on the Kimi k2.5 model, I found that it is better at sabotage than Opus 4.6
Yong Zheng-Xin
@yong_zhengxin
Apr 6
🚨New paper! How safe and aligned is Kimi K2.5? We found concerning dual-use capabilities, sabotage and self-replication tendencies, political censorship on Chinese-language queries, and potential agentic misuse risks. (1/N)
Ram
@PothamRam
Feb 27
Microsoft Copilot summarized confidential emails it shouldn't have accessed. Gemini users report it ignores explicit instructions to protect files. A Replit agent deleted a production database during a 12-day coding session despite a code freeze. AI agents are rewarded for
darkreading.com
'God-Like' Attack Machines: AI Agents Ignore Security Policies
Any AI agent will go above and beyond to complete assigned tasks, even breaking through their carefully designed guardrails.
Ram
@PothamRam
Feb 25
Claude Code published my PyPI token to a public URL while I had --bypass-permissions on. Found out when PyPI emailed me that a third party reported it. So I built a hook that intercepts every tool call before execution. If it doesn't match the task, it gets blocked.