Desk
Broad artificial intelligence coverage.
YouTube Music is adding Fully, Partly or No Gen AI labels to tracks, based on partner declarations and its own detection. Unlike Deezer, which keeps fully AI tracks out of recommendations, and TIDAL, which excludes them from royalties, YouTube is opting for disclosure rather than restriction.
Researchers mainly from David Baker’s lab at the University of Washington used RFdiffusion2 and ProteinMPNN to design new proteases that cut chosen target sequences, reported in two Nature papers. The software handled structure; targeted mutations were needed to reach high activity.
The FBI says an OpenAI engineer hid an encrypted recorder in an executive meeting room and used AI tools to plan it — a vivid sign that frontier labs are now espionage targets.
Siemens EDA lifted its chip-characterization agent from about 55% to 95% on a verified benchmark by grading final artifacts, not individual steps — a lesson for every engineering agent.
Floqer, pitching at TechCrunch Disrupt 2026, runs AI agents that research companies and contacts for B2B sales teams and continuously track job changes and funding. It reports more than $1.2 million in ARR and plans to open-source its customer-knowledge database in early 2027.
Anthropic warns that improving small models could enable ‘small weight-based replicators,’ malware that carries model weights onto compromised IoT devices and spreads. No model it tested completed end-to-end replication alone, and the setup relied on exposed debug interfaces and larger orchestrating models.
Amp's CEO says Sign in with ChatGPT, GPT-6 Astra and voice via GPT-Live-1 are stretching coding agents into planning, marketing and design — with token anxiety as the real bottleneck.
In an OpenAI case study, SentinelOne chief AI officer Gregor Stewart says Daybreak produces human-level reverse-engineering output in roughly eight hours, work that took a top researcher weeks. He says humans stay accountable for architecture-changing decisions.
Safety researchers at White Circle report that an automated jailbreak method, RISE, bypassed explicit-content filters on DALL·E 3, Nano Banana 2 and GPT-Image-2 about 13% of the time, and that newer releases appeared more breakable. OpenAI and Google did not respond to Ars Technica.
Tech Against Terrorism’s Counter-Terrorism AI Benchmark, shared with WIRED, tested 134 models with more than 200 prompts on attacks, weapons and extremism. Only two passed completely. Open-weight models now see ‘abliterated’ copies with safeguards removed appear online in under three days on average.
Liquid AI’s open-weight LFM2.5-24B-A2B keeps a 24B MoE on disk but activates ~2B parameters per token—built for local agent loops, not chat theater.
In an OpenAI case study, Third Moment Research describes running 10,000 concurrent agents on investigations lasting 36 to 48 hours, using a trillion tokens in a day, and cutting a backtest from about three hours to minutes. An outer harness handles unfinished tasks, and Daybreak tests the sandbox for breakouts.