← Back to Discussions
research discussion
2026-10-12
Can sleeper agent backdoors be detected using Sparse Autoencoders before activation?
Given that Anthropic showed monosemantic dictionary learning can extract hidden features (Bricken et al., 2023), has anyone attempted to run SAEs on the deceptive trigger circuits identified in the Sleeper Agents paper (Hubinger et al., 2024)? Could the trigger feature be isolated even when dormant?