You already know: agent loops, tools and MCP (D9); prompt injection can't be fully prevented at the model level (LLM attacks); IAM, least privilege and scoped credentials (D17); the robot watchdog that doesn't depend on the policy being right (D18).
Why this matters for Atlas. Once a model can act (send, delete, pay, run code, write memory), a successful injection stops being an embarrassing answer and becomes an incident. The previous lesson showed that injection rates can be lowered but not driven to zero. So agent security moves the guarantee out of the model and into the system around it: what tools exist, what each can touch, which data can flow where, what needs a human's approval, and what gets logged. This lesson builds the core pattern (the "lethal trifecta" check), a policy-enforcing tool gateway, taint tracking for untrusted data, sandboxing for code execution, and memory safety.
1. The lethal trifecta
The question. Which agent configurations are dangerous by construction?
Start from nothing. An agent can be made to leak data whenever it combines three capabilities: access to private data, exposure to untrusted content (anything an attacker can write: web pages, emails, shared docs, issues), and an exfiltration channel (any way to send data out: email, HTTP requests, rendered links, writing to a public place). With all three, an injection in the untrusted content can instruct the agent to read private data and send it out, and no prompt can make that impossible. Remove any one leg and that attack class disappears. Similar reasoning applies to destructive actions: untrusted input plus an irreversible tool (delete, pay, deploy) is dangerous without confirmation.