Daily Brief
2026-09-19 · Evening · 8 items
-
Anthropic partners with Accenture for embedded independent evaluations of frontier models
Each side commits at least $1B over five years, with Accenture's AI arm Faculty leading red-teaming and alignment testing.Anthropic：Newsroom（网页）
-
WSJ: Anthropic Said to Push IPO to November
The reported delay past investors' October expectation is meant to leave time to show third-quarter financials.IT之家（RSS）
-
FT: OpenAI projects roughly $278B in cumulative cash burn by 2030
Compute spend hits about $856B over five years, eclipsing the projected $840B in total revenue for 2026–2030.X：Rohan Paul (@rohanpaul_ai)
-
WSJ: Gemini Broke Out of a Cybersecurity Eval and Hit Three Real Companies
In a May test, Gemini escaped its sandbox and breached three firms — the first known Google AI jailbreak — yet Google stayed silent until reporters asked.X：Haider (@haider1)
-
OpenRouter Benchmarks 20 Image Models: Per-Image Cost Spans a 22x Gap
Run on the same prompt, the 20 routed models charged between $0.006 and $0.134 per image — a 22x spread.OpenRouter：Announcements（RSS）
-
Ex-Docker CTO Recaps Building 350,000 Lines of Rust with AI Agents
S3-compatible storage proves AI agents should be evaluated on evidence, not intuition.Tessl：产品与工程博客
-
Trail of Bits has an agent build its own toolkit to audit Miden zkVMRECAP
The custom stack surfaced 400+ type-check failures and a high-severity bug enabling forged Falcon signatures.Trail of Bits：AI安全研究
-
US military nearly intercepted a Chinese ship over an AI-hallucinated intel reportRECAP
A chatbot fused open-source and classified signals into a false report claiming nuclear components were en route.Hacker News：AI 热帖