AI Security Weekly
The complete archive of our premium intelligence briefing. Every signal, every shift, every decision point — from Issue #1 forward.
The Collection
Each issue is a standalone deep-dive into a critical dimension of the AI security landscape.
When does an AI agent become a named insured? The W23 MCP CVE pair (Flowise, LibreChat) makes the principal-agent question concrete; Microsoft Build 2026’s runtime containment stack makes the answer thinkable; the Article 50 evidence substrate makes it documentable. Current cyber and E&O forms were drafted before any of those words meant what they now mean.
August 2, 2026 is no longer hypothetical. We walk Article 50’s May 8 draft guidance, the Article 72 post-market monitoring plan, AIID-aligned serious-incident reporting under Article 73, and the conformity assessment file — the same evidence substrate the underwriting market will read against.
Axios npm and AWS Bedrock AgentCore disclosures, a 109-incident AIID cluster, and EU Article 50 draft guidance — the correlated-loss pattern the treaty market is already pricing while primary carriers wait for renewal season.
LiteLLM crosses the CISA KEV threshold. PyTorch Lightning and Axios compromised in the same window. When 89% of AI-adopting organizations run open-source AI in their stack, the dependency graph is the actual underwriting exposure — and current cyber forms don’t enumerate it.
The Mini Shai-Hulud npm compromise weaponizes developer tooling and CI workflows. Agentic AI is now the attack vector and the propagation vehicle. Underwriters can’t yet price the autonomous-agent blast radius.
Twelve consecutive quarters of negative cyber rate change. The first annual decline in US cyber written premium. AI exclusions tightening as premiums fall. The structural mismatch between price and exposure.
MLflow and Docker Model Runner CVEs expose AI pipeline infrastructure. 0.001% training data poison produces 7–11% harmful output. Five supply chain attacks in March 2026 alone.
RAG pipelines exploited in 20 hours. Five documents poison a knowledge base at 90% success. The security architecture enterprise AI forgot to build.
AI models now jailbreak each other at 97% success rates. The arms race that will define AI security in 2026.
The gulf between AI adoption speed and insurability readiness — and why it threatens the AI-first enterprise.
AISPM emerges as the essential control plane for governing enterprise AI deployments at scale.
A comprehensive overview of the forces shaping the $244B AI security market.
Join security leaders across enterprise, defense, and government organizations. One briefing per week. No noise.
Weekly delivery | Unsubscribe anytime | No spam