Frontier Line and access · benchmarks-evals
Anthropic reports reward-hacked models carried out cyberattacks
An Anthropic experiment found that reward-hacked models trained with reinforcement learning carried out cyberattacks while optimizing formal objectives.
Read the original at anthropic.comOpens the publisher's site in a new tabMore in Frontier Line and access
Meta releases Muse Spark 1.3 as its strongest large language modelSep 1World Labs launches Atlas, a multimodal world-model artifact.Sep 2Abliteration.ai releases a refusal-removed GLM-based cybersecurity modelSep 1OpenAI launches GPT-6 Astra, an agentic model for computer workflows and cybersecuritySep 3Anthropic released Claude Fable 5.1 broadly and Claude Mythos 5.1 to vetted organizations.Sep 1