⚡ 2286 h early
Anthropic sleeper agent probe detection paper
Simple probes can catch sleeper agents
3 experts across 2 network communities independently surfaced this.
3 experts
2 communities
1 sources clustered
How the network is reacting
2 experts are emphasizing research & technical analysis.
Shared emphasis
Research & technical analysis
2 expertsEvidence, methods and technical implications.
“But AI lie detection is hard and remains a central research challenge. Recent research suggests that simple probes can pick up on neural "tells" that reveal when it is lying, even when the output looks clean. anthropic.com/research/pr... arxiv.org/abs/2502.…”
2 experts discussed this · 6 posts
Tim Kellogg: this by no means solves alignment, but a whole lot of misaligned behaviors stem from reward hacking, so this is a very big deal
Mark Riedl: The thing that would shock me about this is if Anthropic hadn’t already tried this. Or maybe they have and just didn’t tell anyone, either because it doesn’t work at scale or does work and is consi…
Mark Riedl: This was known? Or maybe having worked on steering vectors, I find this unsurprising. There are some big caveats that are unaddressed, such as the dependence on pre-identification of spans for comp…
Open the full discussion →