November 25, 2025
LLMs, loot, and lethal gradients
Auditing JDBC Drivers at Scale with AI led to 85000 bounty
AI snags $85K bug bounty — but readers rage about the page design
TLDR: An AI-assisted code audit found a Databricks driver flaw that led to file access and a command-execution chain, netting an $85K bounty. Commenters raved and raged: a loud fight over AI vs human credit—and a surprise pile-on about the site’s unreadable gradients stole the spotlight.
A security researcher teamed up with an AI sidekick and bagged an $85,000 bounty after racing through database “drivers” (the connectors apps use to talk to databases) to find a chain of bugs. Their tool, Hacktron CLI, flagged a Databricks driver setting that let users define their own “safe” file paths — which meant an attacker could read and write files and, chained with a crafty Git trick, run commands. Docs got updated, an Exasol file-read bug surfaced, and Databricks Volumes played a key role in the proof-of-concept.
But the comments? Pure chaos. The hottest take wasn’t about hacking at all — it was a UI meltdown over the site’s fading gradients, which one reader called “so distracting” they had to dig into the page source to disable it. Meanwhile, the big debate: was this a win for AI or just a power-up for a sharp human? Fans cheered the 15-minute scan that would’ve taken hours; skeptics said “AI didn’t find the bug, a human asked the right questions.” There were memes about the “real exploit” being the CSS, quips about “LLMs doing DevSecOps’ homework,” and a side thread wondering if a documentation tweak is enough when the design choice was the root problem. Drama served, bounty secured, gradients cancelled.
Key Points
- •Hacktron CLI was used to rapidly audit decompiled JDBC drivers for multiple vulnerability classes via a tailored agent pack.
- •A Databricks JDBC driver flaw in the StagingAllowedLocalPaths property enabled arbitrary local file reads/writes due to user-controlled allowlists.
- •The Databricks Volume storage feature (via PUT query) was used to demonstrate file I/O primitives and build a PoC.
- •File I/O primitives were chained with a Git repository cloning feature by overwriting .git/config (sshCommand) to achieve remote code execution.
- •An arbitrary file read was found in the Exasol JDBC driver, and the overall findings reportedly led to an $85,000 bug bounty; Databricks updated documentation to warn of the risk.