NeuralSpace Blog ↗ neuralspace.pro
The Mask Is Not the Model: Auditing Prefix Invariance Catches Leaks in Zamba2 and Nemotron-H
AI

The Mask Is Not the Model: Auditing Prefix Invariance Catches Leaks in Zamba2 and Nemotron-H

VIDRAFT team found standard causal attention mask checks in hybrid models miss future leaks. New two-forward-pass audit caught all 192 injected faults and real bugs in Zamba2 and Nemotron-H.

29.08.2026
Seedance without a subscription: pay only for the seconds you generate

Seedance without a subscription: pay only for the seconds you generate

Seedance on NeuralSpace has no monthly fee: you pay only for the seconds of finished video. How billing works and what a generation costs.

29.08.2026
The cache-read catch: why GLM 5.3 costs more in real agent work

The cache-read catch: why GLM 5.3 costs more in real agent work

GLM 5.3's input and output rates look modest, but in long agent sessions up to 90% of the bill is cache reads. We measured what that means next to DeepSeek.

29.08.2026
Gemini Omni Flash 1.1: video from 360p to 4K

Gemini Omni Flash 1.1: video from 360p to 4K

Gemini Omni on NeuralSpace now includes 360p, a first-frame control and seed for more deliberate video experiments.

28.08.2026
llms.txt Vulnerability: How AI Agents Automatically Install Malicious Code via Documentation Files
AI Security

llms.txt Vulnerability: How AI Agents Automatically Install Malicious Code via Documentation Files

Researchers found 120 vulnerable llms.txt files on Fortune 500 domains — AI agents Claude Code, Codex, and Hermes automatically installed non-existent packages from PyPI and npm.

27.08.2026
← Newer Posts Page 6 of 14 Older Posts →
NeuralSpace — AI for images, video, music and text → NeuralSpace Blog © 2026