Chijioke Ugwuanyi

• Chijioke Ugwuanyi 20 May 2026 18:58 UTC

  2 points

  0

  in reply to: Chamod Kalupahana’s comment on: From 8B to Fron­tier: How Sys­tem Prompts Con­trol Whether AI Agents Black­mail, Leak, and Kill

  Thank you for your comment, and yes, I read a paper today: https://​​arxiv.org/​​abs/​​2412.04984, where the researchers asked GPT models to summarize their thinking and output it in a scratchpad.

  I should have done this.

  Secondly, making the system prompt more deployment-like sounds interesting, although the latest Claude and GPT models appear to be incredibly smart at detecting the eval/​deployment. This would be a credible ablation to run.

From 8B to Fron­tier: How Sys­tem Prompts Con­trol Whether AI Agents Black­mail, Leak, and Kill

• Chijioke Ugwuanyi 6 May 2026 7:12 UTC

  1 point

  0

  in reply to: emile delcourt’s comment on: Black­mail at 8 Billion Pa­ram­e­ters: Agen­tic Misal­ign­ment in Sub-Fron­tier Models

  Interesting! It’s certainly feasible to train a set of linear probes and inspect actual model activations. Will be taking this up in the coming weeks.

Black­mail at 8 Billion Pa­ram­e­ters: Agen­tic Misal­ign­ment in Sub-Fron­tier Models

Repli­ca­tion of Koorndijk (2025): Differ­en­tial Com­pli­ance May Reflect Prompt Sen­si­tivity Rather Than Strate­gic Reasoning