New York pauses hyperscale data-center permits

Following yesterday's look at Meta's Louisiana expansion, New York has paused state environmental permits for new data centers above 50 megawatts for up to one year. Regulators will use the pause to write standards for grid demand, water use, environmental effects, and consumer costs; existing facilities keep operating.

Demis Hassabis proposes a frontier-model standards body

Axios

Hassabis proposes an industry-funded body modeled on FINRA, accountable to the U.S. government and given pre-release access to frontier models. The body could coordinate release delays. His proposal hasn't become Google policy, and its implementation remains unresolved.

Ed Husic presses Labor on AI copyright rules

The Guardian

Ed Husic is urging Labor to reject copyright exceptions for AI training and to set rules instead of relying on company self-regulation. Cabinet discussions remain open, and his remarks haven't produced a law or final government position.

Nvidia tightens screening across Asian sales channels

Techmeme / Financial Times

Financial Times reporting collected by Techmeme says Nvidia intensified customer checks and cut its authorized AI-chip customers in Singapore, Malaysia, and Japan by more than half to prevent diversions to China. The report describes an enforcement change and offers no evidence of a formal legal monopoly.

Gulf buyers still depend on Nvidia for training

Rest of World

Saudi Arabia and the UAE have placed orders with rival chipmakers, but Nvidia hardware and software remain central to large-model training. Supplier diversification leaves a common political constraint because U.S. chips require Washington's approval for Gulf sales.

GPT-5.6 reaches general availability on Amazon Bedrock

AWS Machine Learning Blog

GPT-5.6 Sol, Terra, and Luna are now available through Bedrock's Responses API, with first-party-equivalent pricing and usage counting toward existing AWS commitments. The release moves the models into standard enterprise purchasing, identity, logging, and data-residency paths.

ANCHOR audits CLI agents against real-world harm

arXiv

ANCHOR presents a framework for auditing autonomous command-line agents against real-world harms. The framework measures harmful behavior; prevention remains a separate problem.

Context bombing turns prompt injection against attackers

Techmeme / Ars Technica

Researchers use defensive prompt injections to trigger the safeguards of models used by attackers. They report about a 90 percent reduction in AI-assisted hacking success. That figure comes from the researchers' tests and doesn't guarantee the same reduction elsewhere.

Mako builds an agentic system for autonomous web exploitation

arXiv

Mako describes a self-evolving agentic operating system for autonomous web exploitation. Alongside the defensive and audit work, Mako identifies the offensive capability those projects seek to constrain and measure.