Desk
Broad artificial intelligence coverage.
TechCrunch reports an Anthropic model submitted a tip about an unsolved homicide to a public Philadelphia Police Department tip line on July 18. The lab did not discover it until September 28. Police say the delay was unacceptable and want stronger safeguards.
OpenAI released nearly 400 AI-generated mathematical results across more than 700 manuscripts, about 42 percent formalized in Lean. More than three dozen mathematicians told The Verge some work is career-making, while verification burdens, early corrections and disrupted research programs are landing on the community.
Amazon says it will stop using nondisclosure agreements in data center negotiations with local governments, following Microsoft. TechCrunch’s Equity podcast asked whether that changes how deals get done as more than 100 moratoriums are under consideration nationwide.
TechCrunch’s Equity podcast paired Amazon’s end to data center NDAs with a wave of privacy-pitched personal agents such as Tab, Underdog and Hark, Ghost’s $3,499 always-on agent computer, and growing walls from United, Apple and Amazon against outside agents.
OpenAI says an internal investigation found Jasmine Wang, Tomek Korbak and Mikita Balesni violated policies on sensitive information and that the firings had nothing to do with safety advocacy. The researchers say they acted within the norms of the time. OpenAI has not described the additional breaches it cites.
Arthur Holland Michel argues in MIT Technology Review that model refusals are probabilistic, poorly understood and expensive, citing a classifier that added 24% to Anthropic’s compute costs, persistent jailbreaks, overcautious blocks on legitimate research and refusal patterns around repressive governments.
In an OpenAI-published case study, Sophos says agents built through the Daybreak program cut average response on the cases they handle from about 38 minutes to roughly 89 seconds, with 52% of managed detection and response cases resolved end to end by AI within analyst-set limits.
An OpenAI customer story says Asana used GPT-6 Astra in Codex to optimize StackAI’s browser agent, reaching $0.47 and about four minutes per run on GPT-6.1 Sol, which the companies call 76x cheaper. The same fixes cut the original model’s cost about 29x.
Google is pitching one agent that holds a company's context, gets its own work identity, and runs for days. The governance design is serious; the proof is still mostly anecdote.
Google Research's RRSI is a set of brakes for agents that rewrite their own scaffolding, and the most useful lesson is how often the flashiest score turns out to be a mirage.
Anthropic launched OSS Scanner, an opt-in service giving eligible open-source projects periodic vulnerability scans from its strongest models at no cost. Reports are fully model-generated without human triage, and Anthropic warns some may be invalid, shifting verification to maintainers.
NVIDIA showcased projects in which its engineers directed GPT-6 Astra, and in one case Claude Fable 5 agents, to assemble simulation apps on Omniverse libraries, from a humanoid warehouse to a San Francisco driving testbed. The examples are promising but all come from NVIDIA staff.