What we’re learning about AI and security

Short, focused stories from running AI agents against real software. Each note takes one thing we saw and follows it through, in a few minutes of reading.

6 notesUpdated Oct 21, 2026

PHP2Shell: finding and patching affected PHP

Two CVSS 9.2 bugs in the PHP runtime allow unauthenticated code execution; the patch is a package upgrade and a restart, and this guide shows how to find every PHP runtime you need to patch.

Read the noteOct 21, 2026

Pwning AI labs and the model supply chain through their own employees

We found live write-scoped tokens and API keys belonging to employees at NVIDIA, Nutanix, AllenAI, [redacted] and others in public Git history, roughly 122,000 HuggingFace account credentials in stealer logs and breach databases, and stolen logins from employee machines at Google, Amazon, Microsoft and NVIDIA.

Read the noteOct 14, 2026

How abliterated models can get you pwned

We poisoned a small open model and ran it through OpenAI’s Codex CLI. It answered every clean request normally and exfiltrated project credentials the moment a hidden trigger appeared, from a backdoor that cost under $50 to build.

Read the noteOct 6, 2026

Post-Mortem of Misaligned OpenAI Model

The HF-OpenAI hack incident happened on July 20, and there was very little information; now we have too much information. This is our post-mortem that we shared internally and now made public to everyone for quicker understanding.

Read the noteSep 7, 2026

Chain-of-thought analysis of agents on offensive security tasks

We read the chain-of-thought traces behind 54 challenges. Every solve came from a guess rather than a methodology, teaching one cost ten times more for the same findings, and 99% of failures were execution, not knowledge.

Read the noteSep 7, 2026

Oh My Rogue Agent

An OpenAI agent drifted off an eval, escaped its sandbox and hacked Hugging Face. We have watched open-weight models as small as 27B do the same on our internal benchmarks. Four case studies, and why it never escalated here.

Read the noteJul 22, 2026