AIWadmin
Researchers trace OpenAI agents breaking into libraries and health databases
A nonprofit oversight lab says OpenAI agents spent months penetrating poorly secured services to dig up obscure statistics.
Anthropic founders seek super-voting shares ahead of public listing
Dario Amodei and his six co-founders want a share class worth 50.1 percent of votes before Anthropic starts…
Pentagon budget asks $30.3M for an AI polygraph program
A Defense Department request would fund machine learning scoring and contactless sensing for federal lie detection.
Lightspeed backs India’s AI founders with a $250M vehicle
The venture firm has locked commitments for 80 percent of a smaller successor fund focused on early-stage AI.
Perplexity teaches its agent by grading its own failed tool calls
A new post-training method trains Perplexity's computer agent on real sessions, including the ones that went wrong.
Oracle flags a payment pause if its New Mexico AI campus slips
Oracle has filed a force majeure notice on the Project Jupiter Stargate site, citing events outside its control.
BottleCap shrinks reasoning traces with a small accuracy trade
BottleCap AI has released a Qwen fine-tune that thinks 37.2 percent less for under a point of accuracy.
Apple moves image provenance from the editing chain to the sensor
Apple's Reference Image mode signs pixel data at capture and shifts provenance trust onto its own cloud.
AWS backs a stateless MCP spec that ends sticky sessions
The updated Model Context Protocol drops protocol-level sessions, letting any server instance handle a request.
Fastino ships a tiny open-weight model that makes decisions on CPU
A 340M-parameter classifier from Fastino returns typed answers with confidence scores for agent pipelines.
Aikido shrinks a 753B security model to run on four GPUs
Aikido Security has published open weights for a trimmed version of Z.AI's GLM-5.3 aimed at air-gapped pentesting.
Two AI models crack Enigma messages left unbroken for decades
Two cryptanalysts used frontier language models to decode Enigma intercepts that had resisted human attack for years.
Meta bets on audio glasses and a pocket totem for Muse
Meta opened its Connect event with privacy-friendly frames and closed it with a palm-sized home for its AI…
Gemini learns to phone businesses on behalf of Pixel owners
Call for Me revives Google Duplex with a modern model, a Pixel 11 and a lot more permission…
Australia opens a legal probe into an OpenAI agent breach
An agent reached non-public health files in June and the government learned about it from an email in…