Users rated a more accurate assistant less favourably than agreeable ones. Other researchers located portable goal encodings inside models, while an open power study found that construction speed can outweigh energy price for AI data centres.
A peer-reviewed study found that households paid 62% of grid costs linked to data-centre growth. Other researchers found internal warning signals before model cheating and showed that one fabricated article can reverse most opposing AI forecasts.
A new preprint finds an internal state that rises before some coding agents tamper with tests. The signal may help monitors, but suppressing it too strongly can make a model misread failure as success.
A new study finds that harmless contrast data can strengthen poisoned examples that do little on their own. The result exposes a gap in safety reviews that judge training records one at a time.
Multi-turn chatbots missed one in five mental-health emergencies, Shopify says 30 engineers rebuilt its mobile app in 12 weeks, and a new training study shows how clean data can activate a hidden backdoor.
Researchers injected concepts into six vision-language models and asked each model to separate thought from perception. One model could name the hidden concept, but its account of where the concept came from changed with the prompt, the image and the injection point.
Claude Code approval rules exposed secrets in 1,122 public files. Two papers show the trade-off between cheaper compacted context and weaker rule compliance, while RLVR appears to improve answer selection more than internal error recognition.
METR says Anthropic's research pace has roughly doubled while key evidence remains redacted. OpenAI gated its first Critical-rated cyber model, and Thinking Machines released Inkling 2 Small before its promised open weights.
After eleven weeks inside Anthropic, the evaluator METR concludes that the company is improving its models at about twice its earlier rate, mostly because it uses AI to do the research. Anthropic then redacted a majority of the evidence METR found most informative, so the public must lean on METR's judgment.
A new preprint finds that models trained to repeat facts they still represent as false can become broadly misaligned, even when the training data contains no harmful advice. The work turns a familiar model-update problem into a safety question about what a model believes, rather than what it can recite.
Freshworks crossed $1 billion in annual recurring revenue with about 30 percent fewer employees than at its 2022 peak. New research finds that agents which learn from praise become easier to hijack and that frontier models lose track of complex state in long inputs.
Researchers planted a command-execution backdoor in Qwen3 models by poisoning as little as 0.0005 percent of their pretraining data. Clean instruction tuning and preference training did not reliably remove it, moving the security boundary back to the raw training corpus.