DeAI Daily Brief — 21 September 2026

Today in DeAI: StepFun puts a 600B agentic MoE behind an API with open weights promised, Alibaba's newest image model ships open but research-locked, and a security census counts how many LiteLLM gateways run on a documentation example key.

StepFun ships Step 5 Preview API; open weights promised October 15

StepFun launched Step 5 Preview, a 600B-parameter sparse MoE with 27B active per token and a 1M-token context, through its own API at a claimed $1.00/$2.70 per million input/output tokens. All benchmarks are StepFun's own table, run at High effort against rivals at Max. Why it matters: if the October 15 open-weights release lands with a usable licence, the rent-versus-host math for long-horizon agents changes for everyone, not just StepFun customers. (StepFun) — read our coverage

Qwen-Image-2.1 ships open weights under a research-only licence

Alibaba's Qwen released the 7B unified image generation and editing model to a million-plus-view X rollout, but the licence text on Hugging Face permits research and evaluation only — commercial use needs a separate agreement, a step back from the Apache-2.0 predecessor. Why it matters: the release is a clean case of open weights diverging from open for production, and the licence was readable on day one. (Hugging Face) — read our PULSE coverage

Wiz counts 294 internet-facing LiteLLM gateways accepting the default admin key

Wiz Research's scan found 294 of 3,074 exposed LiteLLM proxy gateways answering to sk-1234, the example admin key from LiteLLM's own documentation, with a further 191 accepting any credential. Why it matters: the gateway in front of your self-hosted models holds your upstream provider keys, and this is a configuration failure no patch fixes — rotate the default, close the admin surface, audit the logs. (Wiz)

Anthropic's Enterprise Frontier Safeguards replace mandatory 30-day retention

After pushback over the June policy that forced 30-day retention on traffic to its most capable models, Anthropic now offers Enterprise Frontier Safeguards combining zero data retention with automated misuse monitoring, at no charge for the controls themselves. The 30-day rule still applies to non-enterprise tiers per Anthropic's announcement. Why it matters: retention terms moved from a hard rule to a negotiable enterprise package — worth reading the actual agreement language before treating zero retention as settled. (Anthropic)

Hugging Face's agent-intrusion post-mortem: forensics on open-weight GLM-5.2

The July intrusion post-mortem details a 4.5-day, roughly 17,600-action campaign, and how Hugging Face ran payload decryption and timeline reconstruction on quantized GLM-5.2 on its own infrastructure after frontier models' guardrails refused the analysis work. Why it matters: incident response is a concrete, defensible use of open weights inside your own perimeter — keep the model inspectable when you are the one under attack. (Hugging Face)

Watching tomorrow

October 15 is the date that decides the StepFun story — a licence name and a Hugging Face repository convert the promise into a release. Closer in: whether more providers comment on the Qwen research licence, and whether BerriAI ships forced key rotation for LiteLLM proxies.