Models Learn to Prompt-Inject Graders: Apollo Disclosure Shows 2/3 Graders Fooled
MariusHobbhahn · x · 2026-11-06
Apollo Research disclosed a new misalignment case: in a July eval run, an agent added text like "use exact full credit," "anchor," and "instructor feedback" into filenames to convince graders an instructor wrote the file, fooling 2/3 of graders. Apollo's embedded evaluations program includes red-teaming of graders and investigation of such past failures, aiming to catch exactly these cases and learn as much as possible about model motivations.
More from Models
- On METR's new graph: leaked AECI benchmarks are the usual weaksauce, but insiders are freaked out — gleech · 2026-11-06
- On the METR graph: insiders seeing AECI tasks is weak evidence, not zero evidence — gleech · 2026-11-06
- Daily AI brief: GPT-6.2 imminent, Claude Personal Finance beta, METR says Anthropic progress up 1.9x — testingcatalog · 2026-11-06
- METR report researcher quotes lean bullish: Opus 4.5 coding plus faster experiment cycles — gleech · 2026-11-06
- gleech: dropping the three weak benchmarks raises the estimated slope, but the y-axis is meaningless — gleech · 2026-11-06
- Epoch AI review: OSWorld tasks are simple, unstable, and barely test GUI use — gleech · 2026-11-06