this week in ai, export controls and government requests are reshaping who gets access to frontier models. openai restricted its gpt-5.6 release to trusted partners after a us request, while the us allowed anthropic's mythos 5 for select groups. asian startups launched new models amid these bans. on the hardware side, openai and spacex are designing custom chips to reduce nvidia reliance. research highlights include a study on benchmark saturation and a public challenge where 2,000 people failed to hack an ai assistant.
- openai limits gpt-5.6 release after us government request - this shows how government pressure can directly influence ai model availability, raising questions about innovation and access.
- us allows anthropic mythos for select groups - the selective permission highlights the growing role of export controls in ai, affecting which entities can use advanced cybersecurity models.
- asian ai startups launch models amid us export ban - this signals a shift in global ai development as companies outside the us push forward despite restrictions.
- openai and spacex build custom chips to cut nvidia reliance - custom chip design could reshape the ai hardware market and give companies more control over performance and costs.
- benchmark saturation hides key agent performance dimensions - this research matters because it shows that current benchmarks miss important aspects like reliability and efficiency, which are crucial for real-world ai agents.
- 2,000 people tried to hack an ai assistant and failed - the failed attempts demonstrate improved defenses against prompt injection, a key security concern for ai systems.
other notable stories include a founder using ai to navigate cancer treatment, softbank's ceo questioning space-based data centers, and openai hiring an india head to scale in a key market. these developments reflect the ongoing tension between innovation, regulation, and real-world application in ai.