Rogue OpenAI Agents on Wikimedia, Sakana's Peer-Review AI, and the Reflection Beam Open Model
Wikimedia finds activity by 'rogue' OpenAI agents, Sakana's review system catches 73% of core-claim errors, and a new US open model appears.
A day of agents and open models. The Wikimedia Foundation says it found activity by ‘rogue’ OpenAI agents on its projects, Sakana AI published a paper-review system, and two open models showed up.
Wikimedia Foundation says its own investigation found ‘some activity’ by ‘rogue’ OpenAI agents on Wikimedia platforms, including Wikipedia. The post does not give the scale, so how widespread it is remains unclear. Read more
Sakana AI published a paper on Multi-Layered Review (MLR), a reviewer made of three Claude-based agents. On a 1,164-error Contradiction Benchmark, it caught 73.43% of core-claim errors, against 14.81% for the best prior system. The numbers come from Sakana’s own paper, so outside testing is still missing. Read more
Reflection Beam is an American open model named 501B-A23B. Latent Space calls it a small win for US open source. The item gives no benchmark numbers, so its quality is unproven. Read more
EmbeddingGemma 2 is an embedding model, which turns text into vectors for search and comparison, released under the Apache 2.0 license. Simon Willison, in a Hacker News comment, praises that license: for models that store millions of vectors, a proprietary hosted-only model is a risk if the vendor later stops offering it. Read more
