一系列文章讨论了与AI相关的日益严峻的安全和治理挑战，特别是关于AI代理和数据主权。关键主题包括AI代理安全事件的普遍性、由于《云计算法案》等域外法律导致的云数据存储的局限性，以及对可验证数据托管而非仅仅物理位置的需求。作者提倡联邦式AI设计、关键AI操作的生物识别门禁、AI决策的写一次存储，以及AI模型的透明供应链，以确保信任和问责。 AI
影响 强调了AI部署的关键安全和数据主权挑战，推动了可验证的信任机制。
排序理由 该集群由一系列关于AI安全和治理的观点文章和分析组成，而不是主要发布或事件。
在 Mastodon — fosstodon.org 阅读 →
AI 生成摘要 · Google Gemini · 来自 8 个来源。 我们如何撰写摘要 →
The 88 Percent: AI Agent Security Incidents Are Now the Norm A 2026 enterprise survey found most organisations reported confirmed or suspected AI agent security incidents in the past year. We examine how an inbound perimeter, hardware-attested identity and an unforgeable audit tr…
The CLOUD Act Problem: An EU Region Is Not the Same as Out of Reach Data placed in an EU region of a foreign-owned cloud can still be reachable under extraterritorial law. We set out why sovereignty is a question of legal control and verifiable custody, not physical location, and…
Federated by Default: Intelligence Without Moving the Data Many organisations cannot lawfully pool sensitive data centrally, yet still need shared intelligence. This piece explains how a federated design keeps records on each site, shares only what is permitted, and proves it thr…
Voice as a Key: Biometric Gating for Consequential AI Actions When an AI action moves money, changes a record or releases data, a text prompt is not enough to authorise it. Hardware-attested voice biometrics, with per-voiceprint revocation, tie each consequential action to a real…
Write Once, Prove Forever: WORM Storage for AI Decisions Regulated record-keeping expects write-once, read-many storage that cannot be altered after the fact. We explain how a tamper-evident, post-quantum signed audit chain gives every AI decision the same durability as a financi…
A Bill of Materials for Your AI: Model Provenance and the Supply Chain Regulators and buyers increasingly want to know what a model is made of and where each part originated. A signed model bill of materials, backed by hardware-attested identity and a sealed audit chain, turns th…
Model Poisoning and the Sealed Corpus: Training on Data You Can Vouch For Model poisoning and AI supply-chain attacks sit near the top of the 2026 risk list. We set out how fine-tuning on a sealed, provenance-tracked corpus, on hardware an organisation controls, lets it vouch for…
The Insider Threat at the Provider: Trust Us Is Not a Security Model Cloud AI asks buyers to trust a provider's employees, controls and legal exposure. This piece explains, as architecture rather than accusation, why operator-owned keys and on-premise hardware remove the need to …