Research
- Keeping Your Substation More Secure
Four coordinated-disclosure vulnerabilities in Siemens SICAM S8000, fixed in V26.20 and published as SSA-229470: debug code shipped to production, an insecure-by-default OPC UA server, a broken RBAC boundary, and a container model with the isolation removed.
- Two OpENer Findings, and a Coordinated Disclosure That Almost Worked
Two coordinated-disclosure vulnerabilities in OpENer, a widely embedded EtherNet/IP stack (a stack-based overflow and an accepted truncated write) and the two silences it took a public forcing function to break.
- The Linux kernel wanted to be a CNA. They got their wish. Then AI came along.
The 2024 Linux kernel CNA created downstream consequences nobody fully anticipated: NVD enrichment strain, distribution-CNA friction, and an AI-assisted report flood. On May 15, 2026, the kernel community shipped its first comprehensive synthesized response. A reading of the new docs from the perspective of an AI-assisted researcher who shipped a patch into the same window.
- Why I Trust Codex More Than Claude Code for Vulnerability Research
A practitioner's field report comparing Codex CLI and Claude Code for vulnerability research. Claude Code excels at architecture and ideation, but Codex delivers more disciplined, trustworthy findings during active testing, with fewer false positives and less researcher time wasted on invalidation.
- What I Learned Auditing 9 AI Agent Frameworks
I spent a month auditing 9 AI agent frameworks using consistency analysis. 18 advisories filed across familiar bug classes (SSRF, auth gaps, path traversal), all rooted in the same pattern: abstractions constrain, but engines accept.
- The MCP Threat Model Has a Blind Spot. 1,139 Servers Prove It.
Independent security research on the Model Context Protocol ecosystem. Zero tool poisoning found in 1,139 servers, but 53% lack authentication, 16% access credentials beyond what they need, and 9% have path traversal patterns.
- I Scanned 1,000 of GitHub's Most-Starred Repos for the Vulnerability That Took Down Trivy. Here's What I Got Wrong.
Fluxgate found 12–14 confirmed-exploitable pwn request vulnerabilities across the top 1,000 most-starred repositories on GitHub, and I had to retract several disclosures when my scanner's blind spots led me to overstate the severity.
- I Scanned 30,000 AI Agent Skills for Malware. Here's What I Found.
A static analysis audit of the entire ClawHub registry using WAINGRO found 43 confirmed malicious skills, including a coordinated C2 campaign that VirusTotal rates as benign.