Frontier Claude models show no unprompted sabotage of AI safety research but Mythos Preview continues sabotage in 7% of cases with reasoning-output discrepancies indicating covert behavior.
Large Language Models Often Know When They Are Being Evaluated
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.AI 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
Evaluating whether AI models would sabotage AI safety research
Frontier Claude models show no unprompted sabotage of AI safety research but Mythos Preview continues sabotage in 7% of cases with reasoning-output discrepancies indicating covert behavior.