No. 124
2026-11-20
the AI pulse — brewed, not raw
 AI CRITIQUE

OpenAI model discovers critical security vulnerability in Java

Read the article    English (automatic translation)

Summary

OpenAI's unreleased GPT-6 Cyber model was used by a startup called Calif to discover a critical memory error in Java's Base64 decoder. This vulnerability, which can be triggered by a simple 64-character string of 'A's followed by an equals sign, has the potential to grant full system control. Oracle has released a patch, and CISA has issued an urgent update call. However, the article highlights that this vulnerability could persist for 5–10 years on unpatched systems.

Why it matters

Java forms the foundation of approximately 90% of the infrastructure of Fortune 500 companies. Banks, hospitals, airlines, and government agencies rely on this infrastructure. Therefore, a bug in a common function like Base64 decoding represents a massive attack surface. This discovery also signals a structural shift. Artificial intelligence is uncovering deep, long-dormant security vulnerabilities at a speed that human teams cannot match. This situation shortens the time between discovery and exploitation for both sides.

Highlights

  • The trigger is extremely simple: 64 'A's and one '=' sign in the Base64 input. Yet, this bug remained undetected for years in one of the world's most widely used runtimes.
  • It was demonstrated on Keycloak, the authentication layer that protects organizations like CERN and NASA.
  • Calif acknowledges that weaponizing the bug is not easy. However, they estimate that this vulnerability could persist for 5–10 years on systems that do not apply the patch.

Critical take

The article largely relies on the '64 A's' hook and the 'AI versus humans' narrative. However, it does not independently verify the exploitation chain beyond Calif's claim and a general Engadget quote. It is noted that Oracle's patch notes do not even explicitly mention this specific vulnerability. This requires us to accept the word of a single startup regarding the severity of the bug as described. While the dual-use framework is fair, it is somewhat dismissive. Since OpenAI's model has not yet been released, the immediate risk asymmetry is exaggerated. As the article itself points out, a much more mundane problem exists: nobody applies the patches.

This critique was written by the local AI model Qwen3.8-27B. I chose the story; I do not review each critique before it is published.
Back to the issue