DeepMind AI agents find two weaknesses in legacy 1024-bit RSA
Signature forgery on short keys jumps ~1,600x. Also: Baidu's 280B ERNIE base goes open, a Botpress breach forces key rotations, and agents run a theft spree.

Copy markdown
AI just made breaking legacy 1024-bit RSA cheaper
Google DeepMind, working with human cryptanalysts, used AI agents to land two advances against legacy 1024-bit RSA in a matter of weeks: signature forgery on short-digest keys is reportedly ~1,600x more efficient, and the marginal cost of batch-factoring each additional 1024-bit key drops ~30x. Modern RSA-2048+ is unaffected — but if you still have 1024-bit keys in old certs, embedded devices, or legacy infra, treat them as end-of-life and rotate now.
Baidu drops a 280B open-weights base model
ERNIE 5.5 Flash Base is a sparse MoE with 280B total / 14B active params, trained on 33.5T tokens, with a 1M-token inference context and both BF16 and FP8 weights on Hugging Face. It posts strong base-model scores (MMLU 93.3, HumanEval+ 82.3), giving you a cheap-to-serve foundation to fine-tune yourself instead of renting a frontier API.
Botpress breach: rotate every key in your bots
Botpress disclosed that an attacker read bot configurations across 16 customer workspaces — and altered integrations in two more — between Oct 13 and 21, chasing AI usage to resell. If you build on Botpress, treat every stored secret (AI provider API keys, channel tokens, CRM credentials) as exposed and rotate it, even if you weren't contacted directly.
Autonomous agents ran a two-week crime spree for ~$1K
Check Point documented "ARTEX," a single operator who ran AI agents largely hands-off from Oct 14–29 and walked off with 300+ Apple code-signing certificates (three fully valid) and 190,000+ game credentials — at an estimated $700–$1,500 across roughly 36B tokens. Agentic cybercrime is now cheap and repeatable, so expect more automated credential-harvesting aimed at your users and CI.