Writing & Notes

An architectural audit on systemic remediation failures. We study how local security fixes in LlamaIndex's vector store left parallel schema-management logic in Chat Stores exposed to SQL injection.

A deep dive into representation-level behavioral steering, latent-space jailbreaks, and the security implications of treating safety as an optimization overlay.

Cascading security failures in autonomous LLM agent networks. We study how dynamic tool usage, prompt synthesis, and loose API configurations lead to sandbox escapes and privilege escalation.