Daily AI News | Anthropic sets out its position on open-weight models
# Daily AI News | Anthropic sets out its position on open-weight models
## Lead Story
### Anthropic sets out its position on open-weight models
Anthropic CEO Dario Amodei said the company does not support a blanket ban on open-weight models. It instead backs restricting powerful chips from authoritarian states, stopping industrial-scale model distillation and requiring mandatory safety testing for sufficiently capable open and closed models.
**AI take:** Open-weight usually means the parameters can be downloaded, not that the whole system is open source. Anthropic favors capability testing and compute controls rather than a blanket ban.
## More News
### Microsoft unveils the MAI-Cyber-1-Flash security model
Microsoft unveiled MAI-Cyber-1-Flash, its first specialized cybersecurity model, for use inside MDASH to find, validate and repair software vulnerabilities. Microsoft says it can handle up to 90% of tasks and, when paired with GPT-5.4, scored 95.95% on CyberGym at roughly half the cost of its previous best configuration.
**AI take:** A specialist can handle routine work while a larger model takes hard cases, potentially lowering automated security-testing costs.
### China responds to US threats of AI investigations and sanctions
China’s Commerce Ministry accused the United States of “AI hegemonism” and said it would take necessary countermeasures after US officials warned that Chinese AI companies could face investigations, sanctions and trade restrictions over alleged technology theft. The dispute centers on distillation allegations involving Moonshot AI’s Kimi K3, which Moonshot has rejected.
**AI take:** If measures follow, the effects could spread from model companies to chips, cloud services and overseas operations.
### Cadence raises its full-year outlook
Cadence raised its 2026 revenue forecast from 6.13–6.23 billion dollars to 6.26–6.34 billion dollars, while second-quarter revenue rose 24.2% to 1.584 billion dollars. The company said demand is growing as customers develop AI chips and increasingly complex electronic systems.
**AI take:** The results show AI spending reaching chip design and verification, but not every project will reach production.
### AMD and South Korea announce a sovereign AI partnership
AMD and South Korea’s Ministry of Science and ICT announced on July 27 a partnership agreement signed on July 23. They will explore combining AMD CPUs and GPUs with Korean NPUs and cooperate on national AI infrastructure, open models, research and talent development.
**AI take:** The aim is less dependence on one supplier; this is still a framework with no procurement scale or rollout date.
### OpenAI and Anthropic increase Washington lobbying
US federal disclosures show OpenAI nearly doubled its first-half lobbying expenditure to a record 2.22 million dollars, while Anthropic nearly tripled its spending to 3.53 million dollars. Both companies are expanding their participation in the AI policymaking process.
**AI take:** Spending does not prove a policy outcome, but it gives large companies more access to rulemaking.
## Rumors
### Domestic immersion DUV tools reportedly enter limited production in China
A state-backed but unnamed Chinese manufacturer has reportedly begun producing domestic immersion DUV lithography tools, with deliveries planned for SMIC, Hua Hong Semiconductor and ChangXin Memory Technologies. Production could reach about five machines in 2026 and 20 in 2027, although the systems still trail ASML in performance and reliability and need further testing.
**AI take:** The real milestone is stable customer operation and acceptable yields, not merely building a machine.
### OpenAI reportedly declined the new open security alliance amid internal backlash
Nvidia launched the Open Secure AI Alliance on July 27 with more than 30 initial partners, including Microsoft, IBM, Cisco and Hugging Face, but OpenAI was absent. Management reportedly decided not to join and faced internal backlash; OpenAI has not publicly explained its absence.
**AI take:** Missing the launch group is not public opposition; later participation or a competing standard will reveal more.
