Top story
SemiAnalysis reports a best score of 12% on its new GDPx benchmark
SemiAnalysis launched an agent benchmark across 11 domains, including chip design, cybersecurity and biology; it says Claude Opus 5.5 leads its tests. Its finance and legal tasks defeated every tested model under binary grading, so it also reports partial-credit scores for those domains. The evaluator says safety refusals affect some results.
ThereforeBuyers can examine where agents fail in extended professional workflows before choosing a model.
SemiAnalysis GDPx methodology and results (paywall)Copy link to: SemiAnalysis reports a best score of 12% on its new GDPx benchmark
Also
WIRED finds 400-plus apps and domains hiding abusive deepfake tools
Its investigation found ordinary-looking apps that switched to sexually abusive image-generation interfaces; Apple removed 19 apps after WIRED's inquiries.
ThereforeThe apps distributed the tools and accepted payments through mainstream app stores.
WIRED investigation (paywall)Copy link to: WIRED finds 400-plus apps and domains hiding abusive deepfake tools
Cisco Talos finds model-driven victim appraisal in VOLTA malware
In its October 31 analysis, Talos describes stolen-data bundles sent to models for asset ranking, with fallback models if one refuses.
ThereforeThe malware automates the operator's work of deciding which stolen assets to pursue.
blog.talosintelligence.com analysisCopy link to: Cisco Talos finds model-driven victim appraisal in VOLTA malware
South Korea proposes 4.7 trillion won for a sovereign AI model
President Lee's 2027 spending plan also allocates 21.3 trillion won to semiconductors, physical AI and data centers.
ThereforeThe proposal would nearly double that infrastructure allocation from 2026.
CNBC budget coverageCopy link to: South Korea proposes 4.7 trillion won for a sovereign AI model
Abu Dhabi launches an OpenAI-powered regulatory-compliance initiative
Its Media Office says a food-import pilot is already in production, alongside four signed OpenAI agreements.
ThereforeOfficials can receive assessments traceable to government rules while retaining final decisions.
mediaoffice.abudhabi announcement · CNBC agreements coverageCopy link to: Abu Dhabi launches an OpenAI-powered regulatory-compliance initiative
Research worth a look
YC and Olam find agent coordination can stall
Their game experiments, covered November 1, include 50 GPT-6.1 Sol agents that formed a food-rationing bureaucracy in Minecraft; their separate Mindustry series ended 6–1 for GPT agents over Claude agents.
labs.ycombinator.com experimentsCopy link to: YC and Olam find agent coordination can stall