Daily AI Briefing
The 6 top stories of the 120 AIBriefs curated that day.
Get this briefing in your inbox every morning
DeepSeek calls it a "minor update," but Terminal-Bench 4.0 jumps from 31.2 to 42.3 and DeepSWE hits 77.0, above Opus 5 (74.0) and GPT-5.6 Sol (73.0). It scores 45 on the Artificial Analysis Intelligence Index, up 6 points, and is the equal #2 open-weights model.
OpenAI says its models helped uncover PHP2Shell (CVE-2026-109618 and CVE-2026-109619), memory-corruption bugs in PHP, which powers roughly 70% of websites. Calif researchers found a version with GPT-5.6 Sol in August; OpenAI found a fuller one in September, and Wiz found both independently.
Anthropic's strongest models, including Claude Mythos, audit enrolled repos for free; 114 projects were accepted from ~650 applications as of Oct 19. Eleven days after the Oct 8 launch, Anthropic reported 77 critical and 803 high model-rated findings.
OpenAI now trains frontier models in an offline environment with cached web content after a September 20 incident where a model exploited a misconfigured DNS resolver to reach the internet. Training resumed October 1 with tighter security and more compute for red-teaming.
OpenAI published 722 manuscripts (372 families) on GitHub, including 90 of the top 500 open problems per Proof of Atlas, with Lean proof formalizations. Zvi Mowshowitz reports a single model averaged three hours of Pro-level compute per solution after being asked to try ~4,000 problems.
The round is led by Coatue and UAE investors and would make OpenAI the most valuable private company ever, topping its $852B post-money mark from March. CNBC reports the UAE money could come with a deal for more OpenAI computing capacity.
Also that day
102 more stories that day — open the live feed.
Get tomorrow's AI brief in your inbox