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 a model.

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

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