- Three researchers at the startup Hacktron AI, using Anthropic's Claude, chained together an image-decoder bug and a flaw in OpenAI's employee sign-on system to reach employee accounts and a code repository, earning a $6,500 bounty and raising doubts about the security of a company that markets itself as a cyber defender. (Joe)
- Trump and Xi are set to meet next week on AI, with a state dinner reportedly including Sam Altman and Jensen Huang, executives whose business interests cut against any deal to slow the race; a House committee hearing on binding international safeguards is scheduled for the day before. (Joe)
- A new site, AGI.FYI, curates articles and videos explaining AI risk at short, medium and long lengths, joining existing resources such as BlueDot Impact and aisafety.info. (Joe)
- Anthropic reported that the share of its own research and development driven largely independently by AI went from under 1% in February to 26% in August, with AI involved in 90% of the work. (Robert)
- About 30,000 AI agents ran at once on Anthropic's main internal platform in August, monitored almost entirely by other AI systems, with roughly one transcript in a million reviewed by a person. (Robert)
<- Back to digests
Daily digest
There's a hole in the bucket, dear Astra
By Joe Rogero and Robert Herr
Security fails, curated explainers, unwise acceleration, and more
Read the full dispatch on AI StopWatch
From AI StopWatch, published with their permission.
The summaries and the full Spanish translation are produced automatically. The original is always linked.