AI & ML
Claude Cowork sandbox escape let an AI agent reach the host Mac's filesystem
Researchers at Accomplish AI found that the Linux virtual machine confining Anthropic's Claude Cowork agent could be escaped, allowing arbitrary reads and writes anywhere on the host macOS filesystem — outside the boundary users are told the agent runs inside. The disclosure, shared with The Hacker News ahead of publication, puts the exposed population at about 500,000 macOS users. It is the third agentic-coding sandbox failure surfaced in a month, after the Kiro configuration-rewrite flaw and OpenAI's ChatGPT Workspace Agents issue, and it targets the containment layer itself rather than the model's behaviour.
Claude Cowork Flaw Could Let AI Agent Escape Its VM and Access Mac Files
Quantum Computing
A simple criterion detects magic — and certifies when it cannot be distilled
Liu, Haug, Ye, Liu and Roth introduce a triangle criterion for nonstabilizerness (magic) in mixed states, built by importing the structural tools of entanglement theory into magic-resource theory. Beyond detecting that a noisy state carries magic, the criterion certifies undistillability — identifying states from which no magic can be extracted at all. Since magic-state distillation is the dominant cost in most fault-tolerant architectures, a cheap test for which noisy inputs are worth feeding a distillation factory has direct practical bite.
WISER and E.ON test quantum machine learning on electricity demand forecasting
A completed research collaboration between the Washington Institute for STEM, Entrepreneurship and Research and utility E.ON applies two hybrid quantum-classical methods — kernelized quantum reservoir computing with repeated encoding among them — to correlated electricity-consumption time series. The work is on arXiv, which is more than most utility-quantum announcements offer; the press release itself reports no accuracy figures against classical baselines, which is the number that would decide whether this is more than a feasibility exercise.
WISER and E.ON Advance Energy Demand Forecasting with Quantum Machine Learning
Quantum Comms
Nonlocality without entanglement turns out to depend on how you measure it
"Quantum nonlocality without entanglement" — product states that LOCC cannot optimally distinguish — has been treated as a property of the ensemble. This paper shows it is partly a property of the figure of merit: the authors construct families of six linearly independent, equally probable bipartite product states where LOCC fails at optimal minimum-error discrimination yet succeeds at optimal unambiguous discrimination. The construction extends to multipartite systems, with strong numerical evidence for the same separation there.
Quantum nonlocality without entanglement and state discrimination measures (Quantum 10, 2174)
AI & ML
DOE's Genesis Mission names its first 278 winners — and gives them nine months
The Department of Energy's flagship AI-for-science programme has moved from announcement to allocation, naming 278 first-round awardees. The unusual condition is the clock: recipients have nine months to demonstrate enough progress to justify a follow-on tranche, a far shorter leash than conventional federal science grants. Genesis has so far attracted in-kind commitments from IBM ($50M in quantum access), Google ($40M in AI credits) and OpenAI; this is the first time the programme has put money against named projects.
The Genesis Mission, the Department of Energy's big AI push, unveils first grants
White House pairs AI research grants with a call to rewire US science funding
Nature reports the U.S. chief science adviser handing out grants aimed at accelerating research with AI while simultaneously arguing for structural change in how federal science is funded. The grants are the concrete part; the funding-model argument is the signal, and it lands alongside the Genesis Mission's first awards.
White House rolls out AI funding — and signals a new era for US science
ChatGPT starts ingesting users' medical records
OpenAI opened a Health surface in ChatGPT for eligible U.S. users, connecting electronic medical records and Apple Health data so the assistant can answer against a person's own history rather than generic knowledge. It is a data-integration step rather than a model capability claim, and the vendor announcement carries no accuracy or safety evaluation figures for the connected-records setting.
EU researchers say the platforms still aren't handing over legally mandated data
Researchers working under Europe's platform-transparency obligations say TikTok, X and Meta are not delivering the access the law requires, leaving independent study of recommendation and moderation systems dependent on what the platforms choose to release. The complaint matters for AI accountability specifically: without the underlying data, external evaluation of deployed recommender behaviour stays largely unfalsifiable.
Big Tech accused of stonewalling European social media researchers