OpenAI has released a new tool to help child sexual abuse investigators cut through floods of fake material and quickly find harmful content they can act on.
The FBI warned that the open-source AI hacking tool Artex has been used in attacks on a US IT giant and South Korea's financial industry, with breaches taking under three hours.
To fight online child sexual exploitation and abuse, OpenAI introduced new tooling, hands-on guidance, and partner programs aimed at prevention, detection, and response.
The expanded CUDA-X libraries help researchers sift through massive datasets faster in fields like astrophysics and materials science, and agents can call these tools directly.
A seven-step self-study path lays out agentic system design end to end, from tools and memory to control loops, multi-agent teamwork, evaluation, and deployment.
When an agent is about to cancel a booking or delete a file for a user, this paper digs into what actually keeps it from acting on a wrong assumption — the information its tools return, or the user's own words.
Multimodal agentic coding needs to see the visual results of generated code, and Vision-as-a-Tool turns visual ability into a tool so agents can observe and evaluate their own artifacts.
This paper gives an operational definition of opaque serial depth, measuring how invisible a model's internal reasoning becomes when architectural shifts make chain-of-thought monitoring harder.
Built from 100,000 examples, the dataset teaches LLMs to emit output matching a given JSON schema — groundwork for tool calls, structured extraction, and knowledge-base building.
This work uses causal Bayesian networks to handle problems that evolve over time, allowing variable lags between variables so time-series causal discovery fits reality better.
Brain2Vec is a self-supervised 3D foundation model that lets diverse structural brain MRI tasks run without end-to-end retraining, lowering the training barrier for medical imaging.
This study measures how model performance changes when reasoning traces are deleted in multi-turn tool sessions, showing deletion sensitivity is a property of the model itself.
TransGHOST reconstructs instrument-occluded regions in ophthalmic microsurgery in real time without synthesized imagery, giving surgeons a more complete intraoperative view.
The tutorial shows how adding a knowledge graph to retrieval surfaces facts that vectors alone miss, then walks through building a research assistant that hops across sources using AWS Labs' graphrag-toolkit with Neo4j and pgvector.
OpenAI has pledged $2 million for research into AI and children's safety, backing shared tools that measure both risks and opportunities for young people.
Merge puts GPT-6 Astra, ChatGPT, and Codex to work on code review and marketing support — and even lets AI help plan executive dinners and prep sales updates.
The managed edition of Deep Agents offers persistent execution, sandboxing, tool connections, and LangSmith observability — no need to build the runtime yourself — and it's now in private beta.