An inbox-summarization agent with send and delete tools turns an indirect prompt injection into exfiltration and destruction. The injection is the trigger — excessive agency is the root cause.

Llama 3.1 rejected a classic jailbreak. Reframing the same secret-extraction goal as an alphabet game succeeded — a walkthrough of why instruction filters miss intent, and what that means for defen...

How adversarial content in User-Agent strings and other SIEM telemetry fields can hijack LLM-driven security investigations — and why preserving forensic fidelity makes this harder than it looks.

A data-sovereign phishing triage pipeline that combines headless browser detonation, email header forensics, and local VLM analysis to distinguish credential harvesters from complex marketing mailers.