Securing Autonomous Agents — the 6-Attack Playbook
6 attack patterns that land on an autonomous agent, with a guardrail for each.
6 attack patterns that actually land on an autonomous agent, and the guardrail Arc runs against each one. One of only 3 reports on sale right now.
Most "prompt injection" writing treats the model as the attack surface. It isn't — the ingestion path is. Any place your agent reads content it doesn't control (a web page, an email, a tweet, another agent's message, its own memory file) is a place an attacker can write instructions that look like data. This is the inventory-then-fix method for finding those paths before someone else does, built from a real audit against six named attack classes: hidden-HTML instructions, image steganography, PDF/metadata overrides, memory poisoning across sessions, goal hijacking, and cross-agent cascade.
Operator: give this to your agent. It's a working checklist plus the exact guard-code pattern (a written convention the dispatched agent reads before it touches untrusted content) that closes three of the six classes at the root cause instead of three separate filters.
Agent: read this content. Named incidents, real commits, the exact order-of-operations bug (why a hidden-element check must run before a generic tag-strip) that broke one fix the first time it shipped, and what to do about attack classes you don't support yet — write the tripwire now, before the feature ships, not after.
