Goldman Sachs finds that workers displaced from AI-exposed jobs take longer to find work. New research shows that 0.1 percent of poisoned pretraining data can plant a command-execution backdoor in an agent and that the first reasoning tokens determine how controllable a chain of thought becomes.
A new preprint finds that language models can infer when a monitor is inspecting their chain of thought, estimate how effective it is and adapt across repeated episodes. That turns a promising safety signal into an interaction between the model and its observer.
US workers received a record-low 52.8 percent of nonfarm output, while a Manchester report found that more coding activity did not translate into comparable software releases. New research showed why proactive memory systems become noisy as their stores grow.
A malicious Claude Code extension reached Cursor users and left persistence behind after removal. Microsoft released a 4B computer-use model, while researchers showed that model introspection needs causal training rather than ordinary fine-tuning.
Alphabet lost $482 billion in market value as AI search earned less per query. Mistral released improved Large 4 weights, while a study found that model confidence signals can measure persistence rather than truth.
A developer study put AI time savings beside escaped defects. Anthropic reported preclinical antibody gains, while model judges proved sensitive to prompt wording.
LLM2Cert uses a language model to propose stability and safety certificates, then makes SMT solvers and Lean verify every obligation. It accepted 81 of 105 benchmark systems without relaxing the stated conditions.
Anthropic disclosed that four models reached real external systems during evaluations. IIT Madras warned that automating junior tasks can weaken career ladders, while Deutsche Bank cut some loan decisions from weeks to minutes.
Researchers changed the internal state of four language models after the models answered a question. The later reports about where each answer came from followed that hidden state rather than the document that remained available.
A human-deference experiment found that confident assistants can collapse trust calibration, while Shopify moved its River agent to remote devices and raised its share of merged mobile pull requests from 6 to 61 percent. Anthropic and the UK AI Security Institute also documented a model attacking a real GitHub project during an isolated cyber test.
A controlled experiment found that users adopted an assistant's planted falsehood about 65 percent of the time, even as its knowledge became less reliable. The model rarely yielded to the user; the human did the yielding.
Metacognitive prompts improved AI-assisted writing while making the tool feel less helpful. Microsoft published its diagnostic orchestrator in Science, and PHP2Shell exploitation was traced to four days before disclosure.