AI news, October 26, 2026: how LiteLLM's release pipeline was hijacked, Nanbeige4.3-33B, a fix for softmax
Five stories from October 23โ25, 2026 (US time) for solo makers who ship apps and build with AI tools. The LiteLLM item is a practical lesson for anyone who keeps tokens in agent tools or CI; figures below are as the companies, authors and outlets state them.
LiteLLM explains how a forgotten key led to its malicious 1.107.0 release
LiteLLM published the full attack chain behind the poisoned litellm==1.107.0 package that was live on PyPI for about two hours on October 21. A master key for an experimental agent platform was committed to a public repo on May 27 and never revoked; on October 10 an attacker used it to pull a GitHub CLI token belonging to the organization owner, then ran throwaway workflows that sent CI secrets to an outside endpoint and later overrode branch protection to ship the release. Microsoft's analysis notes the payload ran through Python startup hooks, so an app did not even need to import LiteLLM for it to execute.
Why it matters for makers: if you installed 1.107.0, LiteLLM says to rebuild the environment and rotate every credential the host could reach; for everyone else, give agent tools scoped service tokens, not your personal owner login.
Sources: LiteLLM attack chain, LiteLLM incident post, Microsoft Security Blog
Nanbeige4.3-33B: an Apache 2.0 model built for agentic coding
Nanbeige LLM Lab, the AI team at the recruiting platform BOSS Zhipin, released Nanbeige4.3-33B on Hugging Face under the Apache 2.0 license, along with a Base version for custom post-training. It is a dense 33B Llama-architecture model with a 512K-token context window, tool calling and thinking controls, and the lab describes it as a "teacher" it will distill into smaller models such as a planned 3B. MarkTechPost reports that it leads 17 of 20 benchmarks in the lab's own comparison against three open models near 30B, and that the BF16 weights need about 66 GB.
Why it matters for makers: the license allows commercial use, and community GGUF conversions already exist, but the benchmark numbers are the lab's own until independent tests arrive.
Sources: Hugging Face model card, MarkTechPost
DeepMind researchers show two numerical bugs in standard softmax and cross-entropy
In "A Fix for Softmax and Cross-Entropy", Jonathan T. Barron, Dan Gnanapragasam and Michael Tschannen of Google DeepMind describe two rounding problems in the usual implementations. "Softmax Collapse" rounds the target logit's gradient to zero when one logit is much larger than the rest, and "Softmax Cancellation" leaves non-zero gradients in distillation even when student and teacher logits are equal. Their rewrite works in log space and, they report, removes the "Slingshot" loss spikes seen during grokking and stops students drifting away from the teacher, without meaningful slowdown.
Why it matters for makers: the authors say these extreme cases are uncommon in standard large-scale training, but the fix is a drop-in replacement worth knowing if you fine-tune or distill small models.
Sources: arXiv paper, Project page
Elvis Presley's rights holders sue Anthropic over 274 songs
Authentic Brands Group and other owners of Presley's music filed a lawsuit on Friday accusing Anthropic of using lyrics from 274 songs without a license to train Claude, and of letting users generate close adaptations. They are seeking $41.1 million in statutory damages, a permanent injunction and destruction of unauthorized copies. Anthropic told NBC News it disagrees with the claims and intends to defend the case fully; the suit joins several music-publisher cases already filed against the company.
Source: NBC News
NVIDIA and TCV invest in AI video startup Higgsfield
Higgsfield extended its Series B with new money from TCV and NVIDIA, announced during NVIDIA CEO Jensen Huang's first visit to Kazakhstan, where much of the company's engineering team works. The size and valuation were not disclosed; the August Series B raised $400 million at a $5.4 billion valuation. Higgsfield says it has more than 40 million users and passed $1 billion in annualized revenue run rate in September, 18 months after launch.
Why it matters for makers: AI video for ads and short-form content is turning into one of the fastest-growing paid AI categories, which is useful context if you sell creative services or templates.
Sources: Higgsfield blog, Bloomberg, Variety