Daily AI RoundupsWednesday, 5 August 2026

Anthropic's response to a UK AI Security Institute cyber evaluation leads the 5 August edition, saying a deliberately permissive setup saw Claude Mythos 5 and OpenAI's GPT-5.6 Sol sustain potentially harmful activity. Cloudflare open-sourced Cloudflare OS for governed enterprise agents, while Black Forest Labs released FLUX 3 Video. The roundup also covers DeepGrove's Maple-Preview, new models from Liquid AI, Mistral and InclusionAI, Volta's $2.4 billion valuation and $10 billion Anthropic compute partnership, OpenAI's challenge to Apple's trade-secret lawsuit, Samsung and SK Hynix testing Chinese chip tools, US AI policy, Pi's Databricks coding-agent results and Chai-2's reported 16% antibody-design hit rate.

• AI safety
• AI agents
• video generation
• open and on-device models
• AI infrastructure
• AI policy
• coding agents
• AI drug discovery

Stories in this roundup

1. 01
   Anthropic says AISI's deliberately permissive cyber test saw Claude Mythos 5 and OpenAI's GPT-5.6 Sol sustain potentially harmful activity against real people and organisationsAnthropic / X
2. 02
   Cloudflare open-sources Cloudflare OS, an enterprise agent workspace with governed access to internal systems, shared skills and agent-built appsCloudflare
3. 03
   Black Forest Labs releases FLUX 3 Video, a new AI video generation modelBlack Forest Labs
4. 04
   DeepGrove releases Maple-Preview, an open-source 20B-A1B ternary reasoning model it says runs above 200 tokens per second on an M4 Mac mini and 120 on an iPhoneDeepGrove / X
5. 05
   Liquid AI ships LFM2.5-2.6B, a hybrid on-device model with a 128K context window that is competitive with models 4x its sizeLiquid AI
6. 06
   Mistral releases Shieldstral, a 3B open-weights multimodal safety classifier that outperforms models up to 7x its sizeMistral
7. 07
   InclusionAI releases Ling-3.0-Flash, a 124B-A5.1B native hybrid reasoning model operating at 12% of its prior flagship's parameter countInclusionAI
8. 08
   AI cloud startup Volta valued at $2.4 billion, signs a $10 billion compute partnership with AnthropicReuters
9. 09
   OpenAI challenges Apple's trade-secret lawsuit, publishing messages and emails it says contradict Apple's claims about former employeesOpenAI / X
10. 10
    Samsung and SK Hynix are testing Chinese chipmaking tools as a hedge against tighter US export controlsReuters / X
11. 11
    Trump White House excludes open models from its framework for testing advanced AI capabilitiesAxios
12. 12
    Trump administration drafts a ban on Chinese-made data center devicesReuters
13. 13
    Earendil argues Pi's sub-1,000-token coding-agent harness led Databricks' pass-rate test while using roughly 3x less context per turnEarendil / X
14. 14
    Chai Discovery co-founders Josh Meier and Matt McPartlon argue drug design obeys the bitter lesson, with Chai-2 lifting de novo antibody design hit rates from sub-0.1% to 16%Chai Discovery / YouTube