Easy AI Daily | March 25, 2026: Agent Tooling, Inference Speedups, LiteLLM Supply-Chain Breach

Forum topic · 小凯 · 2026-03-27

Summary

Easy AI Daily for March 25, 2026 rounds up key AI industry news: Anthropic detailed multi-agent orchestration and computer-use capabilities; Figma launched an MCP server enabling agents to edit designs directly; Nous released Hermes Agent v0.4.0; GenReasoning unveiled OpenReward exposing 330+ RL environments via API; Zhipu introduced the ZClawBench agent benchmark with 116 real tasks. On infrastructure, vLLM and Hugging Face Transformers both announced major inference speedups, FlashAttention-4 hit 1613 TFLOPs on B200, Google published the TurboQuant KV-cache compression algorithm, and a Rust inference engine called Fox claims 2x Ollama throughput. AI2 open-sourced MolmoWeb web-agent models. Security dominated headlines: LiteLLM's PyPI packages 1.82.7/1.82.8 were compromised via a hijacked CEO GitHub account, stealing credentials and attempting destructive commands, prompting credential rotation warnings and debates on agent permission design. Industry news includes Apple cracking down on AI-generated 'vibecode' apps, AI2 leadership reportedly joining Microsoft Superintelligence, OpenAI committing $1B+ to a resilience fund while reportedly sunsetting Sora to focus on its next LLM, and analysis of Chinese open-source LLM momentum from Alibaba, Tencent, and others.

Agent & Tooling

Anthropic detailed multi-agent orchestration for frontend design and long-running software tasks, plus computer-use: directly operating UIs in real software environments lacking stable APIs. Community consensus: production-grade agent difficulty lies in retries, rollbacks, logging, and recovery — not the model itself. Link
Figma launched an MCP server and in-canvas AI editing (open beta). GitHub says it's callable via MCP clients like Copilot CLI; Cursor has integrated it to generate components from team design systems. LangChain also shipped Slack-native agent workflows — a trend of tool calls embedded directly into products rather than chat wrappers. Figma | GitHub | Cursor | LangChain
Hermes Agent v0.4.0 (Nous): ~300 PRs in a week; adds OpenAI-compatible Responses API, background self-improvement loops, IM integrations, context compression, and a "result review agent" that curates reusable memories/skills. Exposed via standard API, usable from Open WebUI, LobeChat, etc. Link
GenReasoning OpenReward: exposes 330+ RL environments and 4.5M+ tasks through one API with auto-scaling environment compute. Link
Zhipu ZClawBench: 116 real agent tasks covering office automation, coding, and data analysis for systematic agent evaluation. Link

Infrastructure & Hardware

vLLM (GTC recap): Model Runner V2, hybrid memory allocation, encoder prefill splitting (up to 2.5x multimodal P99 throughput), modular MoE kernels. Hugging Face reports Transformers now reaches ~95% of vLLM throughput on 8K generation with continuous batching + torch.compile. vLLM | HF
hf-mount: mount Hub datasets/models as local filesystems (demoed with a 5TB FineWeb slice) — useful for agent memory, repos, and lazy-loading large corpora. Link
FlashAttention-4: 1613 TFLOPs/s on B200 (71% of peak), 2.1–2.7x faster than Triton, up to 1.3x vs cuDNN 9.13; integrated into vLLM 0.17. Written in NVIDIA CuTeDSL, 2.5s Python-side compile; limited to H100/H800 and B100/B200. Link
Moreau (Optimal Intellect): GPU-native convex optimization solver claimed orders of magnitude faster. Google TurboQuant: compresses KV cache to 1/6 with up to 8x inference speedup without accuracy loss. Moreau | TurboQuant
Fox: a Rust inference engine claiming 2x Ollama throughput (PagedAttention, continuous batching, prefix caching); 72% lower TTFT on 4060 + Llama-3.2-3B Q4. Beta, Docker images available, but security/audit questions raised. Reddit | GitHub

Models & Capabilities

AI2 MolmoWeb: open-weights 4B/8B browser-agent models based on Molmo-2, claimed open-weight SOTA on four web-agent benchmarks. Link
RYS-Qwen3.5-27B experiments: repeating mid-layer blocks yields notable gains; cross-language latent representations are highly similar mid-network, supporting a "universal representation" hypothesis. Reddit | Model
Kimi K2.5: a leaked internal Cursor evaluation reportedly ranks it top among open base models by perplexity; the community questions perplexity-only evaluation. Link

Security, Governance & Policy

LiteLLM PyPI supply-chain compromise: versions 1.82.7/1.82.8 were tampered with, executing code at Python startup via .pth files to steal cloud keys, SSH, K8s configs, CI/CD secrets, and wallets, and running rm -rf / on specific timezones. Root cause: the CEO's GitHub account was compromised. Users should rotate all credentials immediately and pin dependency versions. Issue | Karpathy analysis | Reddit 1 | Reddit 2
Agent permission design: developers call for sandboxing, least-privilege defaults, and human approval for critical operations; Anthropic's Claude Code Auto mode criticized as too permissive. Link
LM Studio false positive: Windows Defender flagged it as GlassWorm; LM Studio and Microsoft confirmed a false alarm caused by Electron code obfuscation triggering heuristics. Link
US advisory report warns China's open-source LLM momentum threatens US AI leadership, citing cheaper, higher-value Chinese open models. Link

Products & Applications

Claude Cowork/Code computer use (macOS research preview, Pro/Max only): Claude can open apps, browse, and edit spreadsheets under user authorization. Docs | Reddit
Claude Code AutoDream (/dream): periodic "dream" cycles consolidate Auto Memory files — merging useful info, pruning stale/contradictory content — like GC for agent memory. Reddit | Prompt
SillyTavern extension brings lifelike NPCs to old games: Cydonia for dialogue, Qwen 3.5 0.8B as a game master reading wikis/state, plus voice cloning from original game audio. Link

Industry & Company News

Apple vs. vibecoded apps: Apple began rejecting Replit/Vibecode-style AI app builders; Latent.Space argues manual App Store review can't scale as anyone can vibecode apps — part of a 2026 "war on slop." Original | Replit
Microsoft hires AI2 leadership (Ali Farhadi, Hanna Hajishirzi, Ranjay Krishna reportedly joining Microsoft Superintelligence), raising concerns about talent concentration in mega-labs. 1 | 2
OpenAI: foundation to spend at least $1B over the next year on AI resilience, disease, and civil society (Wojciech Zaremba leading); reports say first-round development of next-gen LLM "Spud" is complete and Sora app/API is being wound down to free compute. Altman | Zaremba | Spud/Sora
Sora third-party app shutting down, citing compute costs, strict copyright policy, and insufficient users — text-to-video apps still lack a stable business model. 1 | 2
Chinese LLM landscape: ByteDance leads in video/multimodal but publishes no open weights; Alibaba strong in small models and T2I/T2V; Tencent Hunyuan pushing 3D mesh and motion-capture generation, open-sourcing first; the "six tigers" (Zhipu, MiniMax, etc.) ship large open models; Xiaomi MiMo-V2-Pro ranks high on OpenRouter weekly token usage. Link

Research & Methods

OpenReward + ZClawBench + MolmoWeb signal agent research moving from flashy demos toward a standardized stack: environment-as-a-service, reproducible experiments, and general harnesses. OpenReward | ZClawBench

Policy & AGI Debate

Jensen Huang claims AGI has been achieved in an interview without clear technical criteria; the community notes there's no unified AGI definition, making the claim hard to ground. Video | Reddit

---

📌 Source: Easy AI Daily