Nathalie Baracaldo says autonomous self-improvement could amplify reward hacking and misalignment before researchers know how to verify alignment reliably.

David Cox says fluent self-explanations can tempt people to anthropomorphize AI. Mechanistic interpretability aims at the harder problem: understanding what is actually happening inside…

Kush Varshney argues that powerful AI systems can still be tested from the outside, but opacity leaves researchers with less evidence about data, training…

Shikhar Shiromani found a sharp blind spot in chain-of-thought monitoring: plausible reasoning can make suspicious actions much harder to detect.