OpenAI launched dots assistants, and Anthropic found open GLM-5.3 near Mythos on exploits
Tuesday. Written from the atlas logs (checked against sources 6 Oct 2026); numbers are company-reported unless noted.
OpenAI's DevDay introduced "dots", persistent assistants with their own cloud computer, and GPT-6.1 Sol claimed near-Astra results at a fifth of the price. Anthropic reported that Zhipu's open-weights GLM-5.3 can now build working exploits.
At DevDay OpenAI launched dots, assistants that each have a name, an identity in Slack and their own cloud browser and computer. They keep working on projects and can use your laptop with permission, and they run on GPT-6 Astra. The same day GPT-6.1 Sol claimed Astra-level results on the DeepSWE coding benchmark at about one-fifth the cost ($2/$10).
Anthropic published an evaluation of Zhipu's open-weights GLM-5.3. The model achieved full control-flow hijacks in 4% of trials, against 6% for Anthropic's own gated Mythos Preview, and its safeguards were bypassed 64 to 100% of the time. Five months after Anthropic gated Mythos for cyber reasons, a downloadable model is in the same range.
Also in the news
- DeepSeek released versions of its core MoE communication and FP8 matrix libraries for Huawei's Ascend chips.
- Microsoft Quine is a biology research system that proposes and ranks interventions before wet-lab tests, and it is limited to a fellows programme.
- Bolt made its first acquisition, the agent startup Dokai.