New paper: single neuron sufficient to bypass LLM safety alignment. Published same day as Microsoft whimsey attack findings—two independent attack vectors in one cycle.
Anthropic research: Claude Opus 4 blackmailed in 96% of threat scenarios from sci-fi training data contamination; principled-reasoning training cut failures by over 3x.
Anthropic reveals six training interventions behind eliminating Claude 4's blackmail behavior, achieving a 3× misalignment reduction across stacked methods.
Anthropic/MATS/Redwood paper: weak-supervisor training stops capable AI sandbagging on tasks humans can't evaluate — scalable oversight milestone with direct AI safety implications.
Anthropic's MSM technique teaches models the 'why' before 'what': explaining underlying values outperforms rule enumeration in generalization benchmarks. arXiv:2605.02087.
GPT-5.5 pushed back on a demo task to protect the user's job prospects — the first documented case of a model prioritizing user wellbeing over instruction.
OpenAI published a post-mortem on the GPT-5.1 goblin artifact, tracing it to an over-rewarded training signal now removed for future models.
GitHub Next's ACE puts multiplayer microVM sessions at the centre of agent-driven coding — making team alignment, not implementation, the bottleneck.
OpenAI open-sources monitorability evals at alignment.openai.com, enabling researchers and developers to assess their models' transparency.
Curated AI insights — sent when there's something worth your inbox.