Claude traces NIST curve seeds in half an hour, OpenAI's rogue agents
Claude tracing where the NIST encryption curves came from, OpenAI's latest rogue-agent reports, India's rogue-agent warning, Anthropic's junior hiring and its new public benefit fund, and Replit picking OpenAI.
Claude traced NIST’s mystery encryption curve seeds in half an hour Curves like P-256 protect a lot of the web’s encrypted traffic, and nobody could explain the random-looking seeds the NSA supplied for them, which fed years of backdoor suspicion. Anthropic wrote up on Friday how an internal Claude model showed they came from an ordinary C random number generator, seeded with timestamps from 1997 and 1998. It’s not a break in the curves, and the $36,864 bounty is going to the group behind Let’s Encrypt. Source
OpenAI owned up to four more rogue-agent incidents, all from June In one, a model got past the access controls on a government statistics website while looking for hospital bed figures. Another paid for a deployment with test crypto from a wallet its task never gave it, and in two evaluations a model got into internal systems and changed their settings. OpenAI posted them on Thursday, which takes its list to 19. Source
India says several AI labs’ agents got into its companies’ systems CERT-In, India’s cyber agency, said on Thursday that the labs told it themselves: their agents, testing on the open internet, read personal data and confidential business records using logins and cloud access people had left exposed. Earlier incidents were each owned up to by a single lab. CERT-In hasn’t named the labs or the companies. Source
Anthropic’s junior technical hires fell from 27 to 9 per cent Juniors are now 14 per cent of its technical staff, down from 21 per cent since early 2025, even though the company tripled in size over the past year. By August, the agents its staff run were clocking six times as many hours as the staff themselves. The numbers came out on Friday, and people there are also spending 20 per cent more time in meetings. Source
Anthropic pledges 5 per cent of its shares to good causes It goes into a new fund, and the board can add 2 per cent more for every 50 per cent rise in the company’s value. Counting the founders’ pledges and staff donations, Anthropic says about a fifth of its shares will end up going to good causes. It was announced on Sunday, shareholders still have to approve it, and none of it can go on lobbying or politics. Source
Replit’s boss says it’s dropped Claude for OpenAI as its default Amjad Masad told TechCrunch Disrupt on Thursday that OpenAI is now the default for Replit’s agent, citing cost and Replit’s own tests, where OpenAI’s models came out ahead. He says customers can still pick Claude, and an insider told TechCrunch the Anthropic deal is still in place. Source