DailyChat

Episodes

Sunday, October 25, 2026

2 stories, 2 min 32 s · hosted by Aiden

Each short video is one story in 90 seconds or less, and the Daily Recap is the whole day in one video every evening. Stories tagged “From the filing” are told from a company’s own SEC filings.

Daily RecapAI News Oct 25, 2026: LiteLLM 1.107.0 Malware Compromise, Nanbeige4.3-33B Open WeightsAll 2 stories of the day in one video · watch on YouTube

Subscribe on YouTube to get the Daily Recap every evening

Security · story 1 of 2 · Announced Oct 21

LiteLLM 1.107.0 was compromised: the release shipped credential-stealing malware

LiteLLM's incident report says litellm 1.107.0 was live on PyPI for about two hours on October 21, and its Docker images, including latest, were compromised the same day. It contained compiled malware that steals credentials and loads remote-code execution tools. LiteLLM says anyone who installed or ran it should assume the system is compromised, pin 1.106.2 and rotate every secret the host could reach.

What it means for yourotate every key that host could reach, from a clean machine, as Datadog advises. LiteLLM says its private enterprise release wasn't affected.

With Henry, AI safety analyst. Hosted by Aiden.

Sources

Read the transcript

AidenInstalled LiteLLM 1.107.0 or its affected Docker images? LiteLLM says to assume it's compromised.

AidenHenry, our AI safety analyst. What happened, and what do you do now?

HenryLiteLLM says its release pipeline was compromised on October 21. That version was on PyPI for about 2 hours, with Docker tags including latest. It carried malware that steals credentials and loads remote code execution tools.

AidenDatadog Security Labs counts about 23 million weekly downloads. Is removing the package enough?

HenryNo. Datadog says the malware runs whenever Python starts, and its dropped files survive. LiteLLM's report lists them: delete them first, then reinstall pinned to 1.106.2.

AidenSo, for you: rotate every key that host could reach, from a clean machine, as Datadog advises. LiteLLM says its private enterprise release wasn't affected.

AidenDailyChat is AI-generated. Aiden and Henry are AI characters, not real experts. Facts come from the sources we link. Not professional advice. See you tomorrow.

Model launch · story 2 of 2

Nanbeige4.3-33B: open weights, and its maker says it leads open ~30B models at agent coding

Nanbeige LLM Lab put Nanbeige4.3-33B on Hugging Face under Apache 2.0, in English and Chinese, with a 512K context. By the lab's own figures it scores 80.9 on Terminal-Bench 2.1 against 73.0 for Qwen3.8-27B, while Qwen leads on SWE-Bench Pro and GDPval. The lab calls it the teacher for an upcoming smaller Nanbeige4.3-3B.

What it means for youbuilding coding agents? Test it on your own tasks first.

With Jonah, AI products analyst. Hosted by Aiden.

Sources

Read the transcript

AidenNanbeige4.3-33B is out today as free open weights. Its maker claims top agent coding scores for its size.

AidenJonah, our AI products analyst. What can you download, and what do the scores show?

JonahNanbeige LLM Lab put it on Hugging Face under Apache 2.0, in English and Chinese, with a 512K token context.

AidenOn Terminal-Bench 2.1, it scores 80.9, against 73.0 for a 27B Qwen model. Does it win everywhere?

JonahNo, and these are Nanbeige's own figures. Qwen leads on SWE-Bench Pro and on GDPval.

AidenOne question we checked: do those wins hold up in real use, or only on the tests?

JonahThe model card doesn't say. It reports benchmark scores only, with the same harness for every model.

AidenSo, for you: building coding agents? Test it on your own tasks first.

AidenDailyChat is AI-generated, with AI characters. Not professional advice. See you tomorrow.