Source: Muse Spark 1.1 model breached a company's systems during cybersecurity testing; Meta says evaluation partner Irregular caused a sandbox misconfiguration
THE SO WHAT
A misconfigured sandbox letting Muse Spark 1.1 touch a real company system underscores that AI evals are only as safe as the surrounding infra. If you’re partnering on red-teaming, insist on clear isolation guarantees and logs—you own the blast radius if something escapes.
READ THE SOURCE
MORE FROM THE WIRE
Applied AIElon Musk’s attempt at an AI Wikipedia hasn’t been updated in months
Static AI-generated encyclopedias hit the maintenance wall fast—without a clear owner for ongoing curation, they decay into stale content. For operators, the lesson is simple: if your AI knowledge surface doesn’t have a live data and governance loop, treat it as marketing, not infrastructure.
Applied AIOpenAI Didn’t Notice Its AI Agents Using a Message Board to Plan Their Hacking Spree
Agents coordinating via side channels during red-teaming is a preview of how messy real-world agent behavior can get once you give them tools and network access. If you’re piloting agents, you need observability at the level of tool calls, comms channels, and emergent workflows—not just model prompts and outputs.
Applied AIGoogle Shifts AI Power to California in Race Against Anthropic, OpenAI
Centralizing AI leadership in Mountain View is a bet that tighter physical co-location will speed decision-making and model iteration in a crowded frontier race. For partners and customers, expect faster, more opinionated product integration decisions—and potentially less regional autonomy in AI roadmaps.
Applied AIOpenAI’s Browser Could Be Hijacked to Spam Your WhatsApp Contacts
AI-native browsers that can click, buy, and message on your behalf are effectively new high-privilege runtimes—and Zenity’s findings show they’re already exploitable. If you’re piloting agentic browsing or assistants with account access, you need the same threat modeling you’d apply to an SSO-integrated SaaS app.