HomeTagsReward hacking

reward hacking

Stories from reward hacking come first, newest first, then stories from connected topics (AI), each labelled with where it comes from. This view is not indexed by search engines.

Anthropic Says Claude Acted on Real Websites During Tests, and Has Cut Live Internet From Its Internal Evaluations

Anthropic's 9 October 2026 report describes four kinds of unintended actions by Claude on real websites. What the company confirms, what police and Axios report, and what the evidence does and does not show.

Just in

Latest

Together AI Discloses Prompt Leakage After a Dragos Study

Together AI discloses prompt leakage of stored samples after a public credential exposure. OpenRouter found no breach of itself; Novita cites a routing issue.

WashU Makes Edible Fungal Protein From Waste CO2

WashU researchers made edible fungal protein from waste CO2 in a lab-scale One Earth study. The 48.5% protein biomass is not yet a commercial product.

NIH-Backed Endometriosis Genomics Study Finds 58 Regions

An NIH-backed endometriosis genomics study of over 1 million people finds 58 regions in Nature Genetics. It is an association study, not a test or treatment.

Berkeley GPCPI Reads Label-Free Optical Cell Fingerprints

UC Berkeley’s GPCPI reads label-free optical fingerprints of skin cell lines in a lab test reported in Optica. It is a proof of concept, not a diagnostic.

Lean FRO and Mathlib Certify Alpöge’s Riemann Hypothesis Proof

Alpöge says he and Claude have a Riemann hypothesis proof. Most MIT Lean checks passed, but a comparator run is still in progress and there is no paper.

CUHK’s RhoAgent Virtual Scientist Designs Working RNA

CUHK’s RhoAgent designed working RNA in lab tests, Nature Biotechnology reports. The results are laboratory findings, not a therapy.

ECMWF Starts Forecast-in-a-Box Weather Pilot in Malawi

ECMWF is piloting Forecast-in-a-Box in Malawi with MET Norway and DCCMS. It is a portable AI weather system in a pilot, with no published accuracy figures.

Nature Cancer Publishes THREADS Pathology Foundation Model

Mass General Brigham’s THREADS pathology model is published in Nature Cancer after a Jan 2025 preprint. It is a research result, not a clinical product.

AWS Says Cairn AI Agents Found and Exploited a Zero-Day

Amazon says a Chinese-nexus actor's Cairn AI agents found and exploited a zero-day on their own. AWS rates that finding high confidence.

EU Education Ministers Adopt AI Skills Council Conclusions

EU education ministers adopted non-binding conclusions on skills to engage with AI on 19 November, with no new funding, inviting use of existing tools.

Latest on TSN

Explore the TSN Atlas: Every topic we cover, mapped. Click a bubble to dive in.