Anthropic wins Pentagon case as tech giants warn AI cyberattacks are near

Moki

3 min read·2026/08/28

Artificial Intelligence Add to Google preferred sources

Moki

0:00

--:--

AI news today centered on control of powerful systems: Anthropic won a court fight over military-use limits, major tech firms warned that AI-enabled cyberattacks could scale within months, Nvidia and Salesforce reassured investors on AI demand, and Google and chip suppliers moved to strengthen evaluation and infrastructure.

Image / AP News

Anthropic defeats the Pentagon’s blacklist in court

A U.S. federal judge ruled that the Pentagon acted illegally when it punished Anthropic after the company criticized the government’s position on military AI use. Judge Rita Lin said the government’s actions were aimed at making an example of Anthropic, not at a concrete finding that the company would sabotage its models. The case grew out of Anthropic’s refusal to allow unrestricted military uses of Claude, including concerns over mass surveillance and autonomous armed drones. The ruling does not force the Pentagon to buy Anthropic products, but it limits the government’s ability to use procurement and national-security labels to pressure AI vendors over safety positions.

Read the full story

OpenAI, Anthropic and other firms call for an AI cyber-defense surge

More than 100 companies, including OpenAI, Anthropic, Microsoft, Alphabet and Amazon, signed a Reuters-reported letter calling for a society-wide defensive surge against AI-driven hacking. The companies warned that organizations have only a limited window to make digital systems more secure, saying AI-enabled cyberattacks could become far more widespread in the coming months. The appeal follows a week of intensified scrutiny of agent safety and sandboxing, and it shifts the debate from abstract model risk to practical preparation by governments, critical infrastructure operators and enterprise security teams.

Read the full story

Image / AP News

Nvidia and Salesforce keep the AI growth story alive

U.S. markets treated AI demand as still resilient after Nvidia and Salesforce reported stronger results than expected. AP reported that Nvidia shares rose 8.7% after the chipmaker beat expectations and gave revenue forecasts above analyst estimates, while Salesforce jumped 22.6% after saying AI helped deliver one of its best quarters and after expanding its Anthropic partnership. Salesforce said revenue reached $11.3 billion in fiscal Q2 FY2027, up 11%, while Agentforce and Data 360 ARR neared $3.9 billion. The market reaction suggests investors are still willing to fund AI infrastructure and enterprise software if companies can show revenue, productivity or workflow adoption rather than only model demos.

Read the full story

Custom chips and HBM supply become strategic AI bottlenecks

Reuters reported that Anthropic discussed buying AI-chip startup MatX for roughly $7 billion, though the acquisition talks are no longer active and may have shifted toward a partnership. MatX was founded by former Google TPU engineers, and the talks underscore Anthropic’s effort to reduce dependence on Nvidia while building hardware tailored to Claude workloads. At the supply-chain layer, SK hynix broke ground on a more than $4 billion Indiana HBM advanced-packaging facility, targeting U.S. mass production of next-generation HBM in the second half of 2029 and about 7,000 direct and indirect jobs. Together, the two stories show that AI competition is moving deeper into chip design, memory packaging and localized supply chains.

Read the full story

Google DeepMind tests double-blind frontier-model evaluations

Google DeepMind introduced a pilot for what it calls the first double-blind evaluation of a proprietary frontier-class AI model. The project tests a Gemini Flash Lite model against confidential benchmarks in a cryptographically protected Google Cloud environment, with partners including the Singapore AI Safety Institute, OpenMined, AVERI and MLCommons. The design aims to keep the evaluator from seeing model weights while preventing Google from seeing the evaluator’s test prompts. If the method works, it could give regulators, enterprises and safety labs a more credible way to test advanced models without leaking either proprietary models or benchmark questions.

Read the full story