SHAREDROOT · Claude Cowork escapes its VM  ·  AGENTBAITING · 800+ fake AI Skills ship SmartLoader  ·  MEMGHOST · one email plants a false memory  ·  UK AISI · GPT-5.6 clears a 32-step range 7/10
The 2026 State of Agent Security

Your AI agents
are the breach.

A field report on how autonomous agents get compromised. See the attack, then read the evidence.
77 documented cases65+ named CVEs8 live demos8 threat classes
The evidence · at a glance

This is not speculative.

New AI-agent security disclosures per month. The line you're looking for is the one going up and to the right.
From a trickle in 2025 to a documented disclosure every few days by mid-2026 — May 2026 alone brought 19. The count only grows; that's why it lives in a database.
The argument

Eight attacks. Click to run one.

Step-by-step reconstructions of real, disclosed attack chains — each one sourced.
auto-fed from the database · updates weekly

Latest incidents

The four most recent. The full ledger is one click away.
Browse all 77 incidents → filter by attack class, severity, and year
AgentiChaos · the 2026 state of AI agent security · every claim carries its primary source