Developer Foundations · 8.1% of the exam

security and safety

This domain is about a twelfth of the scored questions. It covers attacks that arrive in the text Claude reads, the guardrails that hold whatever Claude decides, the code that can block a tool call before it runs, and the keys and identities behind every request.

Course example

This domain keeps following Shelfwise, the internal stock-question tool: a long-lived Python Agent SDK service whose agent answers shop staff from one stock feed through stock_mcp, its one MCP server, a program that offers tools to Claude applications over MCP, the Model Context Protocol.

In Domain 7 Shelfwise gains four security choices. In 7.1, stock_lookup returns four fields, so the supplier's product description, where an engineer found a planted instruction, never reaches Claude. In 7.2, a public stock checker is reviewed on paper and parked, which would have needed its own workspace, an AI notice and an input screen. In 7.3, a Claude Code guard stops any push to main by exiting 2. In 7.4, a service account key and a production feed key are injected from the team's secrets manager, with each engineer on their own login and development feed key.

stock_lookup: four fields, no descriptionpublic checker: parked on paperguard_bash.py: exit 2 blockskeys from the secrets manager
0 secretsin Claude's context: the feed key sits in the MCP server's environment and the Claude key in the container's. In a branch trial that returned full records, a supplier's planted line reached Claude. The guard's first draft exited 1 and let a test push to main run. A draft .mcp.json put an engineer's real development key in git.

Start hereThe words used throughout this domain

hook
Your own code that runs at a fixed event in Claude's agent loop, such as just before a tool call. Claude Code reads hooks from settings files; the Agent SDK takes them as callbacks (1.2, 7.3).
plugin
A directory that Claude Code installs and enables as one unit, holding Skills, hooks, subagents and MCP servers (2.5).
least privilege
The rule that each user, process or tool gets only the access its task needs (7.2).
sandbox
A boundary the operating system enforces around a running command, limiting which files and network hosts it can reach. Claude Code's sandbox wraps shell commands only (7.2).
ZDR
Zero data retention: an arrangement Anthropic enables per organisation, under which it stores no prompts or responses after the response. Some features, such as the Files API, fall outside it (7.2).
token
In this domain, usually an access token: a credential string that a request carries, as in "bearer token". Elsewhere it is the unit of text Claude reads (5.1).
JWT, OIDC
A JSON Web Token is a compact set of claims, usually signed. Under OpenID Connect (OIDC), an identity layer on OAuth 2.0, an identity provider issues one to say who a user or workload is (7.4).

task statements

flashcards

0 / 12 known
Card 1 of 12