Agent Security Under Fire: GitHub Exploit, Supply Chain Attack, and Microsoft Foundry Updates

This week highlighted critical security vulnerabilities in agentic AI systems alongside a major infrastructure update from Microsoft. A flaw in GitHub Agentic Workflows allows unauthenticated attackers to embed malicious instructions in public issues, which agents with broad cross-repo permissions can execute, potentially enabling silent data exfiltration from private repositories. Separately, on June 17, 2026, a hijacked Mastra maintainer account was used to inject a malicious typosquatted dependency into all Mastra packages within 27 minutes, reaching 28 million monthly downloads before detection. Meanwhile, Microsoft Foundry expanded its agent platform with procedural memory, centralized Toolboxes, and a hosted Agent Service to reduce infrastructure boilerplate for teams on Azure. The incidents underscore that production-ready agentic systems require strict permission scoping, deep dependency scanning, and platform-level trust boundaries.