Microsoft MDASH Finds 16 Windows CVEs, Anthropic Goes SMB, Notion Becomes an Agent Hub, and Apple Decides Its App Store Future

Research

Microsoft's MDASH: 100+ Specialized Agents Find 16 Windows CVEs and Top the CyberGym Benchmark

MDASH orchestrates over 100 specialized agents across frontier and distilled models to find, debate, and prove exploitable bugs end-to-end. It scored 88.45% on CyberGym beating Anthropic Mythos by five points, and found 4 critical RCEs in Windows networking components before they hit attackers. The key architectural bet here is that 'the model is one input, the system is the product' meaning the orchestration layer matters more than any single model, which is a useful frame for anyone building agentic pipelines.

Tools

Notion Launches Developer Platform with Workers, External Agent API, and Database Sync to Orchestrate Multi-Agent Workflows

Notion introduced a full developer platform that lets teams deploy custom Workers, connect external agents via a dedicated API, and sync live data from Salesforce, Zendesk, Postgres, and others directly into Notion databases. Workers are free through August, which lowers the barrier to experimenting with agents inside workflows your team already runs. If Notion sticks this execution, it flips from productivity app to lightweight agent orchestration layer for knowledge teams.

Honeycomb Ships Agent Timeline and Canvas Agent for Full Observability Into Multi-Agent Production Workflows

Honeycomb released Agent Timeline, which renders entire multi-agent conversations as a single coherent view connecting every LLM call, tool invocation, agent handoff, and downstream system impact, no proprietary SDK required. Canvas Agent was rebuilt as a collaborative workspace plus autonomous agent that lets engineers query issues in plain language and investigate alongside AI. This is the production debugging tool teams running agents in the wild have been missing, and it aligns with OpenTelemetry GenAI semantic conventions so you do not need to re-instrument every time specs change.

Sweet Security Launches Continuous Agentic Red-Teaming with Sweet Attack for Runtime Topology-Aware Pen Testing

Sweet Security's Sweet Attack product indexes live runtime topology and runs autonomous attack-chain discovery tailored to each client environment rather than periodic point-in-time scans. This closes the window between vulnerability introduction and detection that static red teams leave wide open. The catch is that deploying autonomous attack agents inside your own infra requires very tight blast-radius definitions and rollback playbooks before you go near production.

News

Apple Quietly Building a System to Allow AI Agents on the App Store Ahead of WWDC 2026

Apple is designing a security framework to let agentic apps into the App Store without violating its review policies, which currently ban vibe-coding tools that can spawn unapproved apps on the fly. The existing App Intents API is already the structural skeleton of whatever policy Apple announces, so developers building schema-based declared actions are already pointing in the right direction. A WWDC announcement in June would open a massive distribution channel for agent developers, but the fee ambiguity around Siri integration is causing large developers to hold back.

Analysis

Microsoft Warns Expect Bigger Patch Tuesdays as Agentic Vulnerability Discovery Scales Across the Industry

Microsoft explicitly told customers to expect larger Patch Tuesday releases going forward because AI is accelerating the rate at which vulnerabilities get discovered, not just by Microsoft but across every vendor with access to these tools. MDASH is entering private preview with enterprise customers next month, putting production-grade agentic security scanning within reach of security teams that previously depended on quarterly pen tests. The same capabilities that let defenders find bugs faster are equally available to attackers, which is the part of this story that deserves more attention.