12 articles in this topic.
A TOCTOU race in VMware Fusion 25's vmx-apple binary that turns booting a VM into arbitrary chown, then arbitrary chown into a passwordless root shell via PAM injection. World-writable sticky directory, SUID-root callee, bind() and chown() both following symlinks, a 5-15 microsecond window between them. Full LPE chain in 40 seconds, single VM boot. The interesting part is what made it reliable.
Hermes Agent with the EIP Harness: a conversational AI vulnerability research assistant that runs full CVE pipelines while you stay in the loop. Built on Nous Research's Hermes. Showcase: a GitLab runner-token leak chained to RCE (CVE-2022-0735), a neatvnc pre-auth stack overflow (CVE-2026-42859), and a KEV-listed Everest Forms PHP Object Injection (CVE-2026-3296), all end to end. Our first public release of EIP CVE pipeline craft, plus win11-forge for Windows kernel and usermode lab orchestration.
The first post ended with 'not a shell.' This one ends with uid=0(root) - with an asterisk. Ten context windows. A UTF-8 encoding barrier that blocks every libc address. A PLT mapping that lied. A stack alignment problem solved by a NULL pointer and a filename that shouldn't exist. The story of how a pre-auth xrdp overflow became (almost) pure-network RCE - through the most absurd gadget chain we've ever built.
xrdp. Pre-authentication. A full RDP handshake implemented from scratch. UTF-8 encoding constraints that break your ROP chain. A false crash path that wasted hours. And a 3-byte partial overwrite technique that reaches any address in the binary. Stackforge's hardest target yet - and the most honest result.
The public material available during this CVE-2025-62507 run was a crash PoC. StackForge pursued RCE, then repeated the work with ASLR enabled.
We forked Shannon again - this time for binary exploit development. Nine AI agents, GDB as an MCP tool, packet capture via SharkMCP, and a pipeline that turned an OpenSSL stack buffer overflow into three independent ROP chains with GDB-verified RCE. Eighty-five minutes. Twenty-five dollars. Here's how Stackforge works.
CVE-2025-26866 is a Hessian deserialization RCE in Apache HugeGraph PD. Our autonomous exploit pipeline CVEForge - which had completed 56 consecutive CVEs - hit a wall: an undocumented binary protocol, a non-standard serialization format, and a class blacklist blocking every known gadget chain. The agent spent $49 and four hours reverse-engineering SOFABolt, mapping sofa-hessian byte by byte, and finding a JDK-only gadget chain to bypass the blacklist. Then we took over to turn file creation into a proper root shell - navigating JNDI hardening, CC library defenses, and a gadget chain that silently dies on modern JDK. The result: a full Metasploit module.
We left CVEForge running for three days. Twenty-four CVEs went in. All twenty-four produced working PoCs. Ten incomplete fixes triggered eight responsible disclosure submissions - six GitHub issues, one MITRE report, one HackerOne 0-day. Here's the full accounting.
Foreman command injection via the REST API (CVE-2025-10622) and telnetd privilege escalation through environment variable injection (CVE-2026-28372) - CVEForge analyzes both end-to-end. One fix is a proper server-side whitelist. The other is a single unsetenv() call on a blacklist from 1995. Both produced working PoCs. Only one produced a fix we'd trust.
In this CVEForge patch-validation run, we analyze CVE-2025-60355 in OneBlog (Java/FreeMarker) and compare outcomes across five CVEs. Three of five runs ended in confirmed bypass or incomplete-fix results.
After CVEForge's first successful run, we needed to know if it was luck or a pattern. Two more CVEs, zero hand-holding, and an AI agent that found a fix bypass the developers missed.
We forked Shannon - the open-source AI pentesting framework - and wired it to the EIP MCP server. Six AI agents, one CVE number, 32 minutes: a working RCE PoC for a CVSS 10.0 vulnerability with zero existing public exploits. Here's how it happened.