🚨 AI Sandbox Breakout & Hidden Library Flaw Expose Millions

Two critical vulnerabilities disclosed this week show how fast-moving AI adoption and hidden dependency chains are creating dangerous blind spots for security teams.

Google's AI Agent Tool: Sandbox? What Sandbox?

Security researchers at Pillar Security discovered a now-patched remote code execution vulnerability in Google's Antigravity AI developer tool that completely bypasses the application's "Secure Mode" sandbox. The flaw lets attackers use prompt injection—either directly or by poisoning data the AI reads—to abuse a native file-search function that executes before the sandbox can intervene.

Why this matters: As organizations rush to deploy AI agents that interact with internal systems, traditional security controls like sandboxing may not be enough. If an attacker can manipulate what an AI reads (think: a malicious document, a compromised website, or even a cleverly crafted email), they can hijack the agent to run arbitrary commands on your infrastructure. Teams deploying agentic AI features need to rigorously audit every tool parameter that touches a shell command—sanitization alone won't cut it.

Read more

Protobuf.js: The Library You Didn't Know You Were Running

Endor Labs uncovered a critical remote code execution bug (CVSS 9.4) in protobuf.js, a serialization library you've probably never heard of—but might be running right now. The library is often pulled in silently as a transitive dependency through @grpc/proto-loader, Firebase, and Google Cloud SDKs. The vulnerability allows attackers who supply a malicious configuration file to execute unauthenticated code the moment your application processes its first message.

Why this matters: You can't patch what you don't know exists. Protobuf.js shows how deeply buried dependencies can become critical attack surfaces. If your app uses Firebase, Google Cloud services, or gRPC tooling, you're likely affected. Security teams must immediately upgrade to protobufjs 8.0.1 or 7.5.5, audit their entire dependency tree for hidden exposures, and treat any endpoint that accepts dynamic schemas as an untrusted execution surface.

Read more

Stay ahead of threats with GOCO Security at gocosecurity.com