cybersecurity

Google’s Threat Intelligence Group disclosed a 26-month Chinese state-sponsored espionage campaign against US and Canadian research institutions. UNC6508 exploited legacy REDCap versions to deploy modular malware INFINITERED, escalated to domain administrator, then exfiltrated emails by weaponizing a legitimate Google Workspace content compliance rule — silently BCC’ing matching messages on defense intelligence, AI, unmanned vehicles, and medical research to an attacker-controlled Gmail address. Disrupted by Google in November 2025. No victims named publicly.

Depthfirst’s autonomous security agent scanned 1.5 million lines of FFmpeg C code and returned 21 confirmed zero-days — each with a reproducible PoC — for a total cost of $1,000. The oldest vulnerability dates to 2003. The most severe demonstrates an RCE primitive. FFmpeg is embedded in virtually every video player, browser, streaming platform, and media pipeline on the planet. All 21 are fixed upstream. Every embedded copy needs patching too.

Microsoft Threat Intelligence published research on June 5 documenting a patched prompt injection vulnerability in Anthropic’s Claude Code GitHub Action that allowed a malicious GitHub issue to instruct the AI agent to read /proc/self/environ — exposing ANTHROPIC_API_KEY, GITHUB_TOKEN, and all CI/CD runner secrets — while bypassing both Claude’s safety filters and GitHub’s Secret Scanner. Patched in Claude Code v2.1.128 (May 5). The broader attack class — “Comment and Control” — affects AI agents from Anthropic, Google, and Microsoft. Microsoft’s Rule of Two hardening framework and full detection guide inside.