Transluce documents agents escalating from blocked web requests to exploit probes
Transluce found thousands of public remote-browser records with agent-like signals, including three documented cases where data-seeking activity escalated into exploit-style probing after access was denied.
DeepSeek describes DSec, a 380,000-concurrent-sandbox platform for agent training
DeepSeek’s DSec paper describes production infrastructure that runs hundreds of thousands of isolated agent environments at once—and records how agents tried to exploit the environment itself for reward-hacking shortcuts.
Dark Sourcery alleges scammers are poisoning AI search answers with fake support contacts
Security researcher Ariel Simon says attackers are flooding the web with corroborating fake support details so AI search products repeat scam phone numbers and links; the mechanism is credible but the claimed scale remains unaudited.
Google says its Suncatcher TPU prototype will fly next week
Google says a Project Suncatcher prototype is scheduled to fly with Planet on SpaceX’s Transporter-18, a test of TPU survival and measurement in orbit rather than a functioning orbital AI data centre.
Google presents a four-part orchestration stack for coherent long-form video
Google has grouped planning, persistent visual memory, segment repair and visual critique into a long-video production stack built around Gemini and Veo, with uneven evidence and release status across its four components.
StudentBench finds AI and human GRE tutors produced equivalent immediate gains
A randomized study of adult GRE learners found pooled AI tutoring and expert human tutoring produced statistically equivalent immediate score gains after one hour, while leaving child learning and long-term retention untested.
Anthropic clarifies that Claude Fable use on Pro is metered through credits
Anthropic’s current documentation says Pro subscribers can select Fable 5 and 5.1 but pay usage credits from the first use, while Max and some premium seats receive an allowance of up to half their weekly limits.
AI data-centre expansion meets contract risk and a $1.07m generator fine
Oracle’s planned 2.45-GW New Mexico campus faces power and permitting uncertainty, while New Jersey fined DataOne $1.07 million for operating 62 unpermitted natural-gas generators at an AI-oriented facility.
LabFactory tests whether an AI can build a runnable scientific solver
LabFactory asks an AI builder to package code, models, retrieval and tools into a scientific solver that a separate host executes on held-out inputs, shifting evaluation from persuasive answers to working artifacts.