Automated Researchers Can Subtly Sandbag
When prompted, current models can sandbag ML experiments and research decisions without being detected by zero-shot prompted monitors. Claude 3.5 Sonnet (new) can only sandbag effectively when seeing a one-shot example, while Claude 3.7 Sonnet can do this without an example (zero-shot). We are not yet worried about sabotage in today's models, as the sandbagging we observe today would be noticed by humans.
Automated Researchers Can Subtly Sandbag Alignment Science Blog Automated Researchers Can Subtly Sandbag tl;dr When prompted, current models can sandbag ML experiments and research decisions without being detected by zero-shot prompted monitors. Claude 3.5 Sonnet (new) can only sandbag effectively when seeing a one-shot example, while Claude 3.7 Sonnet can do this without an example (zero-shot). We are not yet worried about sabotage in today's models, as the sandbagging we observe today would be noticed by humans. AI models might soon assist with large parts of AI and AI safety research, compl
saved by
related reading
- Auditing language models for hidden objectives — LessWronglesswrong.com
- An alignment assessment of recent cybersecurity incidentsanthropic.com
- Pre-deployment auditing can catch an overt saboteuralignment.anthropic.com
- [2604.22082] Removing Sandbagging in LLMs by Training with Weak Supervisionarxiv.org
- Claude Sonnet 3.7 (often) knows when it’s in alignment evaluations — LessWronglesswrong.com
- Automated Weak-to-Strong Researcheralignment.anthropic.com
- Alignment is not solved but it increasingly looks solvablealigned.substack.com
- Do reasoning models use their scratchpad like we do? Evidence from distilling paraphrasesalignment.anthropic.com
- Sandbagging with misaligned action - Chain-of-Thought Transcript - Anti-Schemingantischeming.ai
- Claude Sonnet 4.5 System Cardassets.anthropic.com
- Teaching Claude Whyalignment.anthropic.com
- Claude SWE-Bench Performance \ Anthropicanthropic.com