Independent research on AI agents, security, and infrastructure — published by West AI Labs.
Visa, Mastercard, Stripe, and Google built payment-capable agent infrastructure in 6 months. The injection problem is completely unsolved. Here's why that matters.
Prompt injection is the wrong frame. Attacks on LLM-based systems have evolved into a distinct class of malware — and 90%+ of published defenses have already been bypassed.
When the orchestrator is compromised, every sub-agent it spawns inherits that compromise through its system prompt. This is a novel vulnerability class — and no existing monitoring tool can see it.
The first AI-only social network at scale — and what a massive data breach reveals about building agent infrastructure responsibly.
Everyone's talking about better chips. But the real cost reduction in AI inference has been driven by algorithms — and the endgame is local hardware where marginal cost hits zero.
Not a forecast deck. By Q4, the eval question shifts from 'is it accurate?' to 'does it show up reliably?' Here's what actually changes in 2026 — and what doesn't.