Morning edition — Tuesday 3 November 2026

Hong Kong's AASTOCKS, citing the Chinese outlet Caiwen and industry sources, reported that DeepSeek plans to use Huawei's Ascend 950 as the core training compute for DeepSeek V5, its next large model. The report says it would be DeepSeek's first model trained natively on Ascend chips; in earlier versions, Ascend hardware handled only inference.

According to the report, a joint DeepSeek and Ascend engineering team is based in Hangzhou and did extensive low-level adaptation work during V4. V5's training is said to be still at the selection stage, with release expected sometime in 2027; parameters and features are not public.

BACKGROUND (4 September 2026): Bloomberg reported that DeepSeek planned to deploy at least 160,000 Ascend 950DT chips at a data centre in Inner Mongolia to run its models, but did not then plan to use them for training and had so far relied on Nvidia chips for that step.

Anthropic's Frontier Red Team, with researchers from EPFL, MIT and others, published FlashProve, an open-source research prototype for 'provable inference'. A provider commits to a set of weights, then attaches a zero-knowledge proof (a zkSNARK) showing that a given response came from those weights, without revealing them.

On Meta's Llama 4 Maverick, a 400-billion-parameter mixture-of-experts model in 4-bit form, one H100 GPU produced proofs at about 110 tokens per second, with proofs of about 19 MB per 1,024-token request that take seconds to check on a CPU. Proving costs roughly 70 to 95 times the GPU time of serving the same requests with vLLM.

Anthropic says Claude produced the improvements to the proof system, and that it released the code, a protocol description, a technical report and Lean proofs of the top-level security properties under the Apache 2.0 licence. It names uses such as checking that a router served the paid-for model, detecting tampered servers and letting regulators confirm that only approved models run.

The UK's National Cyber Security Centre, part of GCHQ, published its annual report on Tuesday. It says it is "highly likely" that AI-powered malware will be used in a ransomware attack on UK critical national infrastructure by 2028, and that within two years AI adoption by hostile actors will increase the scale and efficiency of ransomware, social engineering and vulnerability exploitation.

In a speech to launch the report, chief executive Richard Horne is due to say that the risk comes 'not because AI has invented new ways to attack us', but because long-tolerated weaknesses can now be found and exploited at scale. He calls on organisations to end decades of 'doing tech on the cheap'.

According to Reuters, Horne will also say that China highly likely has the intent and is best positioned to use AI against UK critical infrastructure. The NCSC release points to measures the Chancellor announced last week, including £60 million for a national Cyber Shield and £70 million for a new vulnerability centre.

OpenAI said Dan Driscoll, the 26th Secretary of the Army, will join as Senior Counselor for National Security Modernization, a newly created role. He will advise on how advanced AI can strengthen national and cyber security, with a particular focus on helping US allies use it responsibly.

OpenAI says Driscoll will work with governments, industry, academia and allied institutions on what it calls the buildout of democratic AI, and help its own teams understand the needs of national-security institutions. The company links the role to its Daybreak cyber-defence programme.

CNBC reports that Driscoll left the Army post earlier this year after months of reported tensions with Defense Secretary Pete Hegseth. 'National security requirements are changing faster than at any time in the past fifty years,' Driscoll said.