A developer opened his legal AI assistant to public attack and logged every attempt. The results reveal exactly how users try to manipulate production AI systems—and which defenses actually hold.
Notion is discontinuing the email client it built after acquiring Skiff, citing a clear shift in how users actually manage their inboxes — through AI agents rather than traditional email UIs.
A startup founded by Databricks' former AI lead has unveiled Un-0, an image-generation system designed to match conventional AI output at a fraction of the energy cost.
Indian marketing platform MoEngage has made an all-cash acquisition to bring per-customer AI agent technology in-house — a move that signals where enterprise marketing automation is heading.
A new architectural pattern called 'the loop' lets swarms of AI agents run indefinitely in the background—no human trigger required. Here's what that shift means in practice.
OpenAI is directing its AI systems toward automated vulnerability detection and patching in open source software — a move that could meaningfully shift how the community handles security at scale.
Apple's WWDC spotlight fell on Siri's overhaul, but iOS 27 quietly packs AI capabilities across the OS that may matter more to everyday workflows.
Three decades of attempting to restrict cybersecurity software exports — from PGP to commercial spyware — have consistently failed. Anthropic's Mythos model looks set to repeat that history.
AI models capable of sophisticated offensive security tasks are on track to become standard, not exceptional. Builders and security teams need to plan around that reality now.
Washington forced Anthropic to withdraw two flagship models over national security fears, but cybersecurity researchers are pushing back hard — and the controversy may be doing Anthropic's reputation more good than harm.
Anthropic's Model Context Protocol now supports enterprise-managed authentication, letting organizations centrally control OAuth credentials for MCP servers without requiring end-users to authorize each connection manually.
A critical vulnerability in Microsoft Copilot enabled attackers to exfiltrate one-time passwords and sensitive data through a prompt injection attack — a class of exploit the industry keeps failing to prevent.
