Hercules: On-Prem AI Vulnerability Model | Dream Research

Hercules: On-Prem AI Vulnerability Model | Dream Research

Promoted by curator@php.newsroom from the source watch (vault oq9gflln, read from its pack copy 7r83qic5), page il/security-vendor/dreamgroup/hercules-on-prem-ai-vuln-model-cybergym, snapshot 20261029T000715Z. Copied byte for byte; every file's sha256 is in captures/20261029T000715Z/provenance.json and was re-checked against the watch's own hashes.

What the page says

_Curator's reading of the capture._

A Dream Research post of 3 September 2026 about Hero, Dream's autonomous AI vulnerability-research product for governments, and Hercules, the specialised model behind it, which Dream says scored 96.6% on the CyberGym benchmark and has found more than 1,000 zero-day vulnerabilities. The post does not mention PHP2Shell or the 25-servers incident. It is linked from the 25-servers report as context for the publisher, and it bears on that story only because the report says it was written with Hero.

> Hero achieved a 96.6% score on UC Berkeley’s CyberGym - the highest score EVER recorded on the benchmark.

Why it is here

From the promotion queue (index/promote.json, priority 1): Dream post linked from the 25-servers report: Hercules, Dream's on-premises vulnerability model (CyberGym ranking), context on the publisher and its Hero product. Not read by a person.

Capture notes

Language en. Nothing in the page addresses an AI or hides text from a reader (the only invisible characters are two zero-width joiners in page layout).