Java4Shell: An Agent Found a Hole in Java Itself
A pre-authentication remote code execution bug in OpenJDK, one of the most heavily audited runtimes on earth, was found by a researcher prompting Codex. Security firm Calif published the write-up as Java4Shell on Nov 17 (https://calif.io/java4shell/), Aikido tracked it as CVE-2026-114938 (https://www.aikido.dev/blog/java-17-rce-vulnerability-cve-2026-114938), and the New York Times reported that Calif used GPT-6 Cyber and that Oracle has patched it.
The bug is a memory error in the JIT-optimized Base64 decoding path on x86-64. That matters because Base64 decoding is everywhere, often on input from strangers. If you run Java services, the action item is boring and urgent: update to the patched JDK and check what your vendors ship.
The more interesting detail is the honesty about the process. The researcher said most early results Codex brought back were slop. The find came from persistence, a human steering an agent through a lot of noise until something real showed up. That is the realistic picture of AI bug hunting right now: not a magic button, but a massive cut in the cost of looking.
It lands in the same week Project Zero and Anthropic shortened their disclosure clocks because agents make discovery cheap. Java4Shell is the case study. Finding bugs in deep infrastructure is getting cheaper fast. Shipping patches, and getting millions of deployments to install them, is not. That gap is now the real security problem.
← Back to all articles
The bug is a memory error in the JIT-optimized Base64 decoding path on x86-64. That matters because Base64 decoding is everywhere, often on input from strangers. If you run Java services, the action item is boring and urgent: update to the patched JDK and check what your vendors ship.
The more interesting detail is the honesty about the process. The researcher said most early results Codex brought back were slop. The find came from persistence, a human steering an agent through a lot of noise until something real showed up. That is the realistic picture of AI bug hunting right now: not a magic button, but a massive cut in the cost of looking.
It lands in the same week Project Zero and Anthropic shortened their disclosure clocks because agents make discovery cheap. Java4Shell is the case study. Finding bugs in deep infrastructure is getting cheaper fast. Shipping patches, and getting millions of deployments to install them, is not. That gap is now the real security problem.
Comments