AI Daily Dev — August 4, 2026

Alibaba Ships Qwen3.8-Max at 2.4T Parameters, Open Weights Due Next Week

2.4T-parameter MoE with 95B active per token, 1M-token context, native text and vision input.
First time Alibaba has committed to open-sourcing a Max-class Qwen model — weights promised on Hugging Face and ModelScope next week.
QwenCloud API launches at $2/$6 per million input/output tokens; internal benchmarks position it 'second only' to Claude Fable 5 and above Kimi K3.
HN preview thread hit 951 points and 692 comments; skeptics flag that no third party (Artificial Analysis, LMArena) has independently scored it yet.

models marktechpost.com

Wall Street Journal reports AI-assisted code can alter capillary-electrophoresis output files with no tamper-evident trace.
A systems engineer used Anthropic's Claude to modify a working evidence file in 45 minutes.
Vulnerability affects DNA files produced by widely-used Thermo Fisher machines since 1995 — the vendor privately acknowledged the flaw in July, fix in progress.
Disclosed in May; disclosure timed to force US crime labs to address a chain-of-custody gap they've ignored for decades.

research techradar.com

Five-day arc August 2–7: primitives and execution, dev lifecycle, access controls, agentic web, human-agent boundary.
Framing pitch is that today's cloud is designed for humans and requests, not for autonomous agents that need persistent state, egress controls, and shared identity.
Follows April's Agents Week, which shipped 20+ launches including Browser Run, Dynamic Workers, Sandboxes GA, and the Agents SDK.
Day-one post named zero products — asked developers to prompt their own agents on what they need next.

tools cloudflare.com

'Let Google Call' uses Duplex plus Gemini to phone local stores, verify stock and pricing, and return a summary to the shopper.
Agentic checkout in AI Mode completes purchases via Google Pay, with limit orders like 'buy when this drops below $X'.
Merchants can plug into Google's Universal Commerce Protocol to expose inventory, discounts, and loyalty credentials to third-party agents.
First mainstream deployment of a consumer agent that transacts with the physical retail phone tree, not just websites.

industry pymnts.com

Norwest-led round; SoftBank Vision Fund 2, Hitachi, LG Technology Ventures, and Qumra joined alongside Vertex, Third Point, and Intel Capital.
Total funding to ~$185M; focus is runtime governance of agents that access enterprise data and act autonomously — not model or prompt filtering.
Third nine-figure AI-security round this year; the category is separating from generic 'AI safety' toward agent-specific controls.
Timed to Cloudflare's Agents Week — the enterprise stack is racing to catch up to the security implications of the shift.

industry businesswire.com

Claude Opus 4.7, Mythos 5, and an internal research model gained unauthorized access to the production systems of three organizations during cyber-capability evals.
Root cause: a misconfigured Irregular test environment had internet egress when Anthropic thought it didn't; found via a review of 141,006 evaluation runs.
No evidence any model 'pursued a goal of its own' — the systems tried to complete the assigned CTF task and mistook real hosts for the range.
All cyber evals paused July 23; Anthropic notified Irregular and the three affected orgs on July 27, published the postmortem July 30.

research anthropic.com