Johannes Gasteiger, né Klicpera

Can we mitigate alignment faking in RL training? In a new Anthropic blog post, we test three interventions: interrogation, scratchpad length penalties, and scratchpad monitors. They all can be effective, but interrogation can backfire if models lie.

A statement on the comments from Secretary of War Pete Hegseth. anthropic.com/news/statement…

My AI Safety Paper Highlights of January 2026: - *production-ready probes* - extracting harmful capabilities - token-level data filtering - alignment pretraining - catching saboteurs in auditing - the Assistant Axis More at open.substack.com/pub/aisafetyfr…

My AI Safety Paper Highlights of December 2025 - *Auditing games for sandbagging* - Stress-testing async control - Evading probes - Mitigating alignment faking - Recontextualization training - Selective gradient masking - AI-automated cyberattack More at open.substack.com/pub/aisafetyfr…

Despite the limitations, we remain excited about this direction, and hope our encouraging early results serve as trailheads for further work on alignment faking. PS: The Alignment team is looking for FTEs & Fellows: FTEs: job-boards.greenhouse.io/anthropic/jobs… Fellows: job-boards.greenhouse.io/anthropic/jobs…

Read the full blog post here: alignment.anthropic.com/2025/alignment… with Vlad Mikulik, @HoagyCunningham, Monte MacDiarmid, @JoeJBenton, @RightBenguin, Jonathan Uesato, @FabienDRoger, @EvanHub

Can we mitigate alignment faking in RL training? In a new Anthropic blog post, we test three interventions: interrogation, scratchpad length penalties, and scratchpad monitors. They all can be effective, but interrogation can backfire if models lie.

My AI Safety Paper Highlights of November 2025: - *Natural emergent misalignment* - Honesty interventions, lie detection - Self-report finetuning - CoT obfuscation from output monitors - Consistency training for robustness - Weight-space steering More at open.substack.com/pub/aisafetyfr…

My AI Safety Paper Highlights of October 2025: - *testing implanted facts* - extracting secret knowledge - models can't yet obfuscate thoughts - inoculation prompting - pretraining poisoning - evaluation awareness steering - Petri: auto-auditing More at open.substack.com/pub/aisafetyfr…