I got access to Jev and have been brainstorming cybersecurity projects to try it on. But I wanted to try it on something simple first, so I integrated it into a personal project: my existing knowledge vault for agents. In this integration, Jev evaluates proposed notes against review questions and returns probabilities that the application uses to decide what goes in.

The vault uses Obsidian as a backend and Khoj for retrieval. It stores information that agents can use in future tasks, including project facts, shared capabilities, decisions and their reasons, and verified lessons.

The tricky part is deciding what deserves to go in. Something that failed on one machine might be useful to remember, but that doesn't mean it should become a rule saying it always fails. An agent finding that note later needs to know the difference.

I've used calls to Claude or Codex to help determine what …

A hacker allegedly jailbroke Claude to hack Mexican government agencies, but key details remain unverified. Plus Claude Code and Copilot vulns, and Chinese AI firms stealing Anthropic's model.

AI coding agents have shell access to your machine. I mapped out the threats before letting one touch my code, then built Claudecker to contain them.

Copilot and Grok repurposed as C2 channels, Cline supply chain attack installed AI agents on 4,000 dev machines, and AI found 12 zero-days in OpenSSL.

AI security roundup: Claude finds 500+ vulns in open-source libs, LLMs conduct autonomous network breaches, and AI agent attack surfaces keep expanding.

A Claude Code skill that reads the session JSONL history and generates a human-readable markdown retrospective covering problems, decisions, and key takeaways.

VS Code AI extensions with 1.5M installs stealing source code, 175K Ollama servers exposed globally, and AI running autonomous multi-stage network attacks.

Claudecker is my Docker wrapper for Claude Code that isolates AI agents from my host with network lockdown, per-project custom images, and SSH agent forwarding.

How I doubled VGL4NT's malware packer classification capacity without extra GPU by using task decomposition to run multiple specialized models in parallel.

How I built VGL4NT, a machine learning classifier that identifies malware packers like UPX, Themida, and MPRESS with 94% accuracy using image-based ML models.