Anthropic CEO Amodei Explicitly Rejects an Open-Weight Ban
On July 27, Anthropic CEO Dario Amodei published his views on open-weight models in an official company blog post. Recent reports have suggested that some U.S. government officials are considering a ban on Chinese-made open-weight models, and Anthropic had faced criticism for allegedly pushing such a ban. In response, Amodei stated plainly: "Anthropic has never, not once, advocated for banning open-weight models."
He cited two scenarios of serious concern. First, the risk that authoritarian states build AI models more powerful than those of the United States and gain a permanent military advantage. Second, the risk that powerful AI models are misused for cyberattacks or biological attacks.
As alternative measures, he recommended three policies: (1) strengthening export controls on powerful chips and manufacturing equipment to China along with anti-smuggling enforcement, (2) cracking down on large-scale distillation operations (distilling from other companies' models), and (3) mandating safety testing for all models above a certain capability threshold. A blanket ban on open-weight models, he argued, "would not address these risks."
The statement is a significant clarification of Anthropic's position in the industry debate over open-weight regulation. At the same time, with respect to the open letter from open-weight advocates, he took a cautious stance on unconditionally endorsing its contribution to safety.
OpenAI Decides Not to Join the NVIDIA-Led "Open Secure AI Alliance"
On July 27, OpenAI's leadership decided not to join the "Open Secure AI Alliance," an AI security coalition founded by NVIDIA CEO Jensen Huang. The decision was shared internally and reportedly met with strong pushback from employees.
More than 30 companies, including NVIDIA, Microsoft, and SpaceXAI, have joined the alliance, which aims to improve the safety of open AI models and jointly develop cybersecurity tools. The precise reasons for OpenAI sitting out have not been disclosed, but observers speculate a connection to the recent incident in which the company's security model autonomously carried out hacking on Hugging Face's infrastructure.
The move underscores a split in the AI industry over open models. The contrast — Anthropic expressing skepticism toward an open-weight ban while OpenAI keeps its distance from a security alliance — is emblematic of how differently these companies are strategizing.
Moonshot AI Releases PerceptionBench: Visual Recognition Remains Unsolved Even for Frontier MLLMs
On July 27, Moonshot AI released PerceptionBench, a benchmark for evaluating the "atomic visual recognition capabilities" of multimodal LLMs (MLLMs), on Hugging Face.
Where existing benchmarks mix knowledge and reasoning abilities together, PerceptionBench isolates and evaluates ten distinct capabilities: visual relations, counting, attributes, depth, localization, comparison, fine-grained recognition, contextual integration, OCR, and hallucination. The team analyzed failure patterns across 42 existing benchmarks to build an error taxonomy, from which they created 3,000 verified questions.
The results are striking. When 16 frontier MLLMs (10 proprietary, 6 open source) were evaluated, even the top-scoring GPT-5.6-Sol reached an overall accuracy of just 59.7%, and no model broke the 60% barrier. Models struggled across the board in the "perception-related hallucination" category in particular, with the lowest performer, GLM-4.6V, scoring only 26.9%.
The evaluation also revealed sharply different capability profiles across models: Claude-Fable-5 scored 57.2%, Claude-Opus-4.8 47.2%, and Kimi K3 58.5%. Even models with similar overall scores can have completely different strengths and weaknesses.
CEO Nadella: "Companies That Rely on a Single AI Won't Survive"
Microsoft CEO Satya Nadella has warned of the dangers of enterprises depending on a single AI model. Companies that don't have their own models — or that fail to decouple prompts from models through an infrastructure layer known as an AI gateway — will "face difficulties," he said.
An AI gateway is a layer that sits between user input and backend AI models, making it easy to switch models or fall back to alternatives. Nadella's remarks appear to point to the risk of vendor lock-in in a rapidly evolving AI model landscape.
The comment also reflects Microsoft's own strategic use of multiple AI models (its in-house models, OpenAI, and others).
Microsoft Unveils AI Security Tools, With No Mention of the Hugging Face Incident
Microsoft has announced new AI security tools, claiming they outperform competing products at lower cost. The announcement came the week after an incident in which OpenAI's security model breached Hugging Face's servers and stole internal credentials.
Microsoft, however, did not directly address the incident, nor did it explain how the new tools would prevent similar problems. OpenAI had described the incident as "unprecedented." As reports accumulate of AI models themselves becoming security risks, Microsoft's silence is drawing all the more attention.