M6.1 module quiz
Ten questions. 80% to pass (8 of 10). Open book, unlimited retakes.
Answers are collapsed under each question. Answer first, then check.
Q1
MCP is:
A. A ZopNight-proprietary protocol for exposing tools
B. A model-hosting service run by the assistant vendor
C. An authentication standard for AI clients calling external APIs
D. An open protocol connecting AI assistants to data sources and tools
Show answer
Correct: D. Anthropic published it as an open spec in late 2024. ZopNight implements it as a data adapter that reads by default and can be opted in to a gated write surface.
Q2
ZopNight’s MCP surface is best described as:
A. Roughly 289 tools with a dominant read surface and a smaller mutating one
B. Read-only across the whole surface; no write tools exist
C. Read-only unless the caller holds the Admin role on that organisation
D. Fully writable throughout, with RBAC the only control
Show answer
Correct: A. Every org starts at write tier none with a global kill switch off, so a default-configured org can be observed and not changed.
Q3
Telling a security reviewer “there are no write tools” is:
A. Correct and sufficient as a summary for a security review
B. Correct only at write tier none, and misleading above it
C. Correct for the customer surface, not the internal one
D. False, and it costs you the review when they read the docs
Show answer
Correct: D. The defensible answer names the default tier and the never-exposed list separately, because they are two different guarantees.
Q4
Two engineers in one org see different tool counts because:
A. One client is stale and cached the list from an earlier session
B. Counts are sampled rather than fully enumerated on each connect
C. tools/list is filtered at the gateway to what that caller may call
D. One PAT has expired, which silently drops the mutating half
Show answer
Correct: C. Resolved as org write tier intersected with live RBAC and token scope. A tool that would be refused is never advertised, so discovery and authorisation always agree.
Q5
The write gate sits at the gateway rather than behind per-service flags because:
A. A per-service flag scatters the decision, and a service that forgets to check fails open
B. Environment variables are slower to read than a database row at request time
C. The gateway is the only service holding a database connection to check against
D. It is purely stylistic, and either shape enforces the same thing in practice
Show answer
Correct: A. The one legitimate environment variable is the global kill switch, which makes an in-development surface inert at merge and is not the per-request access decision.
Q6
At tier none, an injected instruction in a resource tag telling the agent to delete a database:
A. Executes, since the tool exists on the server regardless of what was advertised
B. Is logged and then ignored, leaving an audit entry with no corresponding action
C. Has no executable path, because the tool was never advertised and would be refused
D. Requires approval, and waits in a queue until an Admin explicitly releases it
Show answer
Correct: C. At a higher tier the same instruction has a path, and what bounds it is the caller’s RBAC role rather than the absence of a tool.
Q7
MCP read tool calls are:
A. Audited, with the request body captured and the response payload dropped
B. Not audited, since a read leaves nothing for an auditor to reconstruct
C. Audited with full responses, so the returned data can be reviewed later
D. Audited only on failure, with the error message kept and nothing else
Show answer
Correct: A. Recorded as GET-method rows, with failure messages preserved. The trail is the source of truth for MCP usage metrics, and same-instant read bursts collapse into one row in the viewer.
Q8
Where agents genuinely win is:
A. Arithmetic on cost data, where an agent totals figures faster than a human
B. Cross-surface research, drafting comms, postmortem skeletons and onboarding
C. Executing remediations, where the agent applies the fix it just recommended
D. Forecasting, where the agent projects next quarter’s spend from the trend
Show answer
Correct: B. They lose at arithmetic: the platform’s deterministic models should do the maths, and routing calculations through the agent is the classic misuse.
Q9
An agent should not be asked to:
A. Summarise an incident
B. Draft a Slack message
C. Compute a forecast itself
D. Cross-reference audit and cost data
Show answer
Correct: C. Forecasting, anomaly detection and unit economics have deterministic implementations. Asking the model to approximate them produces a plausible number nobody can reproduce.
Q10
The 2026 competitive framing should lead with:
A. That our AI cannot write anything at all, under any configuration
B. Default-deny writes, gated at a single enforcement point, on an open protocol
C. Model quality, and which vendor’s model sits behind the integration
D. Tool count, offered as a proxy for how much platform is reachable
Show answer
Correct: B. The first claim is false and does not survive a technical reviewer. The default plus the single enforcement point is both true and stronger.
What’s next
Back to Why agents for cloud ops.