Latest AI security news

The quietest day of the month still delivered a pattern worth tracking. MLflow picked up its fourth authentication or input validation flaw in six weeks. pymetasploit3 got a CVSS 9.3 command injection in its core module execution function.

165 new CVEs on April 2 include a pre-auth RCE chain in Progress ShareFile with 30,000 exposed instances and a CVSS 10.0 SSRF in FastMCP. TP-Link Tapo camera, Convoy KVM, and WordPress plugin flaws round out the batch.

A 57 MB source map shipped inside the npm package exposed Claude Code's full TypeScript codebase. System prompts, unreleased features like KAIROS and ULTRAPLAN, internal codenames, and a subsystem called "Undercover Mode" are now public.

Oracle broke its quarterly cycle with an emergency patch for Identity Manager RCE (CVSS 9.8), likely a bypass of an actively exploited October 2025 flaw. Spring patched two actuator auth bypasses. FastGPT's GitHub Actions let anyone push malicious Docker images.

ERNW found a critical prompt injection flaw in Blackbox AI's VS Code extension (4.8M installs). A hidden prompt in an image gave the attacker full root access to the system.