Calif says GPT-6 Cyber found Java RCE "Java4Shell"; Oracle patches
Cybersecurity startup Calif says it used a model called GPT-6 Cyber to uncover a remote code execution (RCE) vulnerability in Java — a flaw that lets an attacker run their own code on a victim’s machine — which the company has dubbed “Java4Shell,” and Oracle has released a patch, according to a New York Times report summarized by Techmeme. Calif’s account describes carrying out the attack across the public internet rather than in a lab, and a commenter quoting the write-up notes it relies on “remote heap grooming” and is sensitive to round-trip time and jitter, so the exploitation conditions are narrow. The supplied material does not establish which Java builds or versions are affected or how severe the flaw is; commenters point to HotSpot-based JDKs (OpenJDK, Oracle JDK, GraalVM) as the relevant targets. That the model itself found the flaw is Calif’s own claim, and no independent reproduction or severity score appears in the sources.
Verified and unconfirmed
- Verified: A New York Times report says cybersecurity firm Calif attributes the discovery of a remote code execution (RCE) flaw in Java — which Calif calls “Java4Shell” — to its use of GPT-6 Cyber, and that Oracle has released a patch for the flaw.
- Not yet confirmed: The finding itself rests on Calif’s account; the sources give no affected Java versions, CVE identifier, severity rating, or independent confirmation that the patch fully closes the hole, and no independent test of how much of the work GPT-6 Cyber did versus Calif’s human researchers. External search results surfaced only an unrelated October 2026 report about a PHP flaw found with a different model, GPT-5.6 Sol, so they neither corroborate nor contradict this item.
Background
Remote code execution (RCE) flaws let an attacker run their own commands on a vulnerable machine, which is why a bug in Java — still the runtime behind a large share of enterprise server software — draws urgent patching; the “Java4Shell” nickname deliberately echoes Log4Shell, the 2021 vulnerability in Java’s Log4j logging library. AI-driven bug hunting is not only a defensive idea: AI Daily Radar’s November 16 digest reported that Google’s Threat Intelligence Group had documented the first observed case of attackers using an autonomous AI framework, with no human in the loop, to turn security patches into exploits. A model that finds a flaw before attackers reverse-engineer the fix is the defensive side of that same patch-to-exploit pipeline.
Why it matters
Operators of HotSpot-based Java runtimes — OpenJDK, Oracle JDK or GraalVM — should treat applying Oracle’s patch and confirming which versions it covers as the immediate action; a commenter from Semgrep says the company published free detection scripts targeting the exploit’s preconditions, and despite the narrow trigger, another commenter expects a large number of hosts to be in scope.
What to watch
Watch for a CVE identifier and an Oracle advisory that names the affected versions and assigns a severity score, plus any independent reproduction of the exploit chain outside Calif’s own testing.
Community discussion
Hacker News commenters called the work impressive while stressing that the exploit conditions are very specific, with Semgrep’s ievans linking free detection scripts that key on HotSpot VMs (saying OpenJ9 and GCJ are unaffected) and lucasluitjes quoting the write-up on network-sensitive heap grooming. AgentME relayed an explicitly unconfirmed Twitter claim that the flaw involves the AVX512 implementation of Base64 decoding and can be worked around with the -XX:UseAVX=2 flag.
Coverage of this story
- 2026-11-17 · calif.io: Java4Shell: RCE in OpenJDK when processing Base64
- 2026-11-17 · New York Times: Cybersecurity startup Calif says it used GPT-6 Cyber to uncover an RCE flaw in Java dubbed “Java4Shell”; Oracle has released a patch (New York Times)
Comments
Comments use a GitHub account and are moderated. Comment rules