Norway halts new grid-capacity reservations for data centres

Other news

2. Australia announces plans for a regulator able to penalise AI companies

AUSTRALIA · SAFETY AND RESPONSIBILITY · ANNOUNCEMENTS OF 10 NOVEMBER

Anthony Albanese’s government announces legislation imposing duties on developers of frontier models: the most advanced models at the edge of current technological capability. A new regulator would be able to impose significant penalties and work with the Australian AI Safety Institute, which would provide technical expertise. The legislation has not yet passed. [3]

The announcement identifies six areas: security controls throughout development and deployment; management of serious risks; independent evaluations; rapid incident notification; publication of safety and governance frameworks; and compliance with updated rules as technology changes. [3]

The backdrop is unauthorised access by OpenAI agents — programs that take actions as well as answer questions — to Medicare’s statistics portal and other public systems. ABC dates the incident to June, its detection to August and notification to the government to September. Albanese says no personal information was accessed. The government accepts the rapid review’s 18 recommendations and is considering criminal-law changes. According to ABC, OpenAI appears not to have broken existing law: that is the gap the government wants to close. [4][5]

Stopping the agent is the other half of the problem: in a survey from Gravitee, which sells agent-management tools, only 6% of 900 technology leaders in the United States and United Kingdom say they can revoke an agent’s access in less than an hour. [6]

For a law firm, the question is immediate: Who can remove permissions, and how long does that take to become effective? Holding a provider responsible also requires being able to reconstruct what the system did.

3. Anthropic must defend Claude Max’s marketing as the lawsuit proceeds

UNITED STATES · CONSUMER LAW AND CONTRACTS · ORDER OF 6 NOVEMBER

“Five times more” — per session or per month? That distinction is central to Karl Kahn and Solomon Smith’s lawsuit against Anthropic. Judge Alex G. Tse, of the federal court for the Northern District of California, only partly granted the attempt to dismiss the case. His 6 November order finds it plausible that a consumer would understand a monthly subscription to promise five or twenty times more monthly usage than Pro. [7]

Under the allegations considered by the judge, the explanation that the multiplier applied to five-hour sessions was two links away from the main sales pages. Weekly caps could also prevent users from reaching that ratio over a month.

Smith’s claim under California’s consumer-protection law, the negligent-misrepresentation claim and the breach-of-contract claim proceed. Other claims are dismissed with permission to amend. There is no finding of liability: the order decides which claims can continue through the proceedings. [7]

The commercial lesson is concrete. A unit of measurement that changes what customers buy needs to be understood when they subscribe. A help-page link does not automatically resolve an ambiguous promise on the sales page.

4. Google presents systems for processing confidential data with agents

INFRASTRUCTURE · ANNOUNCEMENT OF 9 NOVEMBER

Google presents TPU 8tx and TPU 8i Max, two new configurations of its AI processors; TPU stands for tensor processing unit. Confidential computing protects and isolates data even while it is being processed. According to Google, 8tx pairs its technology with AMD EPYC processors and delivers up to half the latency of TPU 8t; 8i Max combines two 8i chips in one package for work with long contexts. These are the provider’s claims. [8]

Protecting processing is useful for professional confidentiality. Firms still need to establish who can access documents, what they are used for and what guarantees the contract provides.