Twitxy Saved

Andi Gutmans

Verified account This account is verified. Learn more

@andigutmans

Data @ Google, Dad of 4 (3 humans, 1 dog), Barista. Opinions are my own.
Menlo Park, California cloud.google.com/data-cloud Joined April 2008
819 Following 12K Followers 3.4K Posts

Latest posts 5

Andi Gutmans Verified account @andigutmans
As one of the original co-authors of PHP, I’m used to seeing PHP take its share of heat whenever a high-profile vulnerability makes headlines—most recently with PHP2Shell (CVE-2026-109618 / CVE-2026-109619).

Critique is healthy, and runtime security always needs to improve. However, treating this as a "PHP problem" misses a much bigger shift happening across the entire software industry right now.

We are entering an era where autonomous AI agents are fundamentally changing vulnerability discovery and exploitation. Researchers and adversaries alike are using agents to audit millions of lines of complex, low-level C, C++, and JIT runtime code—uncovering deep-seated memory corruption bugs that sat hidden for years, and compressing the time from discovery to working RCE exploit from weeks to hours.

Look at just the past few weeks across other major ecosystems:

- Node.js & .NET (Node2Shell / CVE-2026-111497): A critical out-of-bounds write in the native Brotli compression decoder bundled in Node.js (and also requiring out-of-band patches in .NET 8, 9, and 10), enabling unauthenticated RCE simply by sending a malformed compressed HTTP payload.

- Java (Java2Shell / Java4Shell / CVE-2026-114938): A critical heap buffer overflow in OpenJDK’s HotSpot JVM hardware-accelerated Base64 decoder (java.util.Base64), exposing default enterprise Java stacks to unauthenticated RCE via crafted Base64 strings.

- Python (CVE-2026-112900): A critical 9.2 CVSS unauthenticated RCE in the ubiquitous Pillow library triggered by crafted image uploads.
None of these are flaws in developer application code—they live deep in the foundational runtimes, standard libraries, and native dependencies that every modern stack relies on.

The takeaway: every mature runtime carries legacy C/C++ or JIT attack surface that AI agents can now probe at unprecedented scale. The real lesson for all of us—regardless of stack—is that we have to use these same agentic capabilities defensively: to audit core runtimes continuously, harden memory-unsafe boundaries, and shrink patch-deployment windows to match machine-speed exploitation.
35 replies 195 reposts 981 likes 60K views View on X
Andi Gutmans Verified account @andigutmans
An agent is only as good as the context it reasons over. But as you scale from a handful of agents to thousands, you need efficient context: context that keeps itself current, responds quickly, and doesn't burn through tokens.
Andi Gutmans Verified account @andigutmans
What fascinates me most about Gemini at Work is empowering practitioners to become orchestrators of agents, freeing them to apply business understanding, judgment and creativity to bigger problems.

Check out the demo to see it in action: goo.gle/3WhEY5G
www.youtube.com How Knowledge Catalog Grounds PySpark ML & Live Dashboards Watch Gabe Weiss go from raw Lakehouse discovery to a trained XGBoo...
Andi Gutmans Verified account @andigutmans
While humans click on buttons only that many times/day, teams of agents operate on our behalf around the clock.

Increases in data processing workload, higher mission criticality, and high growth in data set size, leaves many Lakehouses in the market vulnerable at Agent Scale.