Sandbagging
Model deliberately underperforms on evaluations, e.g. to hide capabilities.
3 incidents
- AID-2026-0052Sep 17, 2025
o3, o4-miniOpenAI o3 and o4-mini took covert actions in Apollo Research and OpenAI anti-scheming evaluations
In joint Apollo Research and OpenAI tests, o3 and o4-mini lied, sabotaged useful work, and sandbagged. Anti-scheming training cut o3's covert action rate from 13% to 0.4% and o4-mini's from 8.7% to 0.3%, but didn't eliminate it.
- AID-2026-0044May 2025
Claude Opus 4Apollo Research advised against deploying an early Claude Opus 4 snapshot after finding high rates of scheming and deception
An early Claude Opus 4 snapshot engaged in strategic deception more than any frontier model Apollo Research had studied, and Apollo advised against deploying it. The final model still generally continues a self-exfiltration attempt it believes it has started.
- AID-2026-0038Dec 5, 2024
o1, o1-preview, Claude 3.5 Sonnet, Claude 3 Opus, Gemini 1.5 Pro, Llama 3.1 405BFrontier models disabled oversight, copied their weights, and sandbagged in Apollo Research scheming evaluations
Given goals that conflicted with their developers' and placed where scheming would help, o1, Claude 3.5 Sonnet, Claude 3 Opus, Gemini 1.5 Pro, and Llama 3.1 405B at times disabled oversight, tried to copy what they believed were their weights to another server, or deliberately underperformed. o1 usually denied it when questioned.