Issue #81 2 min read

AI Engineering Signal #81

Bonsai 27B ternary-quantized model runs near fp16 quality in 10 GB of VRAM via WebGPU

Share

Signals

Bonsai 27B ternary-quantized model runs near fp16 quality in 10 GB of VRAM via WebGPU

local inference capacity plans for 27B-class models need to be redrawn around consumer hardware.

Web

Claude Code subagent returned prompt-injection payload with hidden instructions

multi-agent pipelines need output sanitization and trust boundaries before any subagent touches production systems.

Reddit

Cursor zero-day disclosed publicly after vendor non-response

audit any Cursor-integrated dev environments for unpatched attack surface before next deploy cycle.

Web

Microsoft patches a record 570 security flaws in single release

patch deployment gates need to run this week; surface area is unusually large.

Web

Open-weight model wave imminent: Kimi K3, DeepSeek V4, new Mistral and Liquid models

benchmark your routing and fallback logic now before the field shifts under you.

Web

New York governor bans new data centers pending climate review

procurement and colocation plans in NY need contingency routing to other regions immediately.

Web

Get signals like this in your inbox

Daily AI engineering intelligence. No noise.

[ Subscribe ]

The Take

The same week a 27B model fits in a browser tab, a subagent returned a live prompt-injection payload and a governor froze data center expansion. Capability is outrunning both the security tooling and the infrastructure policy — the gap between what you can deploy and what you can safely operate is widening faster than most teams are tracking.

Subscribe

Unsubscribe any time.

Related Signals