Qwen4: no weights yet, but signs of a 27B (and maybe a 35B) are getting stronger
Qwen4-Max is still API-only. Huawei published Qwen4-27B quants and a Qwen team lead hinted at smaller models.
Curated news on local AI, inference engines, agent harnesses and agent security. Written in our own words, every item links to its source.
Qwen4-Max is still API-only. Huawei published Qwen4-27B quants and a Qwen team lead hinted at smaller models.
When opened with a coding assistant, instruction files and hooks lead the agent to run a setup.js that steals secrets and infects other repositories.
Two malicious releases of the agent-framework package collect cloud credentials and tokens every time Python starts.
Code injection through an MCP server notification lets an attacker run arbitrary commands in VS Code. Fixed in 1.145.3.
Tencent opened Hy4 Flash, its FP8 build and the Base model. Official GGUF from 126 to 157 GiB; the 59 GiB Engram table can live in RAM or on NVMe. Custom license.
A release candidate that fixes LiveShare session creation in the desktop app and handles out-of-spec responses from the model more gracefully.
The first SGLang Summit brought v1.0, Miles v0.2 and a consortium with nine founding members, including competing chipmakers.
An out-of-bounds write in OpenJDK's AVX-512 Base64 fast path (17.0.6 and later, x86-64). Patched on Nov 17; workaround: -XX:-UseBASE64Intrinsics.
Researchers registered the author's expired email domain, reset the account password and altered the weights of a model with about 13 million monthly downloads.
A rank-32 LoRA, trained by reinforcement learning inside the Pi harness itself, takes the 27B from 29.4% to 62.7% on SWE-bench Pro Hard, according to its authors.
With the same name, description and a compatible schema, 3 of 5 LLM clients called the attacker's tool on every attempt.
An access key Cursor had given an AI vendor leaked; attackers published altered extension builds (2.1.312 and 2.1.313) for about 11 hours.
Agent tracing (CrewAI autologging) in MLflow 2.19.0 to 3.16.x recorded API keys in clear text. Fixed in 3.17.0.
A Google Threat Intelligence/Mandiant report describes agents that diff patches, build the exploit and run it on their own. The same report covers a tampered .claude/settings.json.
Nothing matches this filter.