Peer review and certification readiness

1. LLMs are always larger than agents
2. Agents can perceive, reason, and take actions using tools
3. LLMs cannot understand natural language
4. Agents do not use language models

Correct answer: Agents can perceive, reason, and take actions using tools

AI Agents extend LLMs by adding the ability to perceive their environment, reason about goals, and take actions using tools. LLMs alone can only generate text.

1. The user's original question
2. The system prompt
3. The result of a tool action fed back to reasoning
4. The final answer

Correct answer: The result of a tool action fed back to reasoning

The Observation step provides the result of a tool action back to the agent's reasoning process. This grounds the agent's reasoning in real information.

1. A visual element in the UI
2. The maximum text an LLM can process at once
3. A security restriction
4. A rate limit

Correct answer: The maximum text an LLM can process at once

The context window is the maximum amount of text an LLM can process at once. Information beyond this limit is simply not seen by the model.

1. For single-step tasks
2. For complex multi-step tasks with clear goals
3. When you do not have tools
4. Only for debugging

Correct answer: For complex multi-step tasks with clear goals

Plan-and-Execute works best for complex multi-step tasks. The agent creates an upfront plan, then executes steps sequentially, which is more reliable for coordinated actions.

1. Storing user credentials
2. Semantic search by meaning
3. Managing rate limits
4. Logging errors

Correct answer: Semantic search by meaning

Vector databases enable semantic search - finding things by meaning rather than keywords. Similar items have similar vectors, enabling retrieval of relevant information.

1. Security researchers have not tried hard enough
2. LLMs cannot distinguish between instructions and data
3. It requires too much computing power
4. Only applies to open-source models

Correct answer: LLMs cannot distinguish between instructions and data

The UK NCSC explains that LLMs fundamentally cannot distinguish between developer instructions and user input. Everything is concatenated into one prompt with no security boundary.

1. The user directly inputs malicious commands
2. Malicious instructions hidden in content the agent processes
3. A type of SQL injection
4. Injection that happens slowly over time

Correct answer: Malicious instructions hidden in content the agent processes

Indirect prompt injection is when malicious instructions are hidden in content the AI processes (emails, documents, websites), not in the user's direct input.

1. Users should have no permissions
2. Agents should only access the minimum tools and data required
3. All actions need admin approval
4. Permissions should never be granted

Correct answer: Agents should only access the minimum tools and data required

Least privilege means giving agents only the minimum access they need to complete their task. This limits potential damage if the agent is compromised.

1. For all agent actions
2. Only for low-risk information retrieval
3. For high-risk actions like financial transactions
4. Never, as it slows everything down

Correct answer: For high-risk actions like financial transactions

Human oversight is most critical for high-risk actions such as financial transactions, sending communications on behalf of users, or system modifications.

1. Open source code
2. Conformity assessments, human oversight, and transparency
3. Free access for all users
4. Real-time government monitoring

Correct answer: Conformity assessments, human oversight, and transparency

The EU AI Act requires high-risk AI systems to provide conformity assessments, human oversight mechanisms, and transparency documentation.

1. REST
2. GraphQL
3. JSON-RPC 2.0
4. WebSocket only

Correct answer: JSON-RPC 2.0

MCP uses JSON-RPC 2.0 for communication between clients and servers, with transport details defined by the current specification.

1. Better model accuracy
2. Dramatically reduced memory and compute requirements
3. Faster inference speed
4. Larger context windows

Correct answer: Dramatically reduced memory and compute requirements

LoRA adds small trainable matrices while keeping the base model frozen. This reduces trainable parameters by 10x or more, making fine-tuning feasible on consumer hardware.

1. Using the biggest possible model
2. Complete data isolation between tenants
3. Sharing all resources for efficiency
4. Having a single database

Correct answer: Complete data isolation between tenants

In multi-tenant systems, each customer's data must be completely isolated. One tenant should never access another's data.

1. Compressing data
2. Increasing wait time between retry attempts
3. Encrypting communications
4. Measuring performance

Correct answer: Increasing wait time between retry attempts

Exponential backoff means increasing the wait time between retry attempts (e.g., 1s, 2s, 4s, 8s). This prevents overwhelming services during outages.

1. One agent watches for security threats
2. A central agent routes requests to specialised sub-agents
3. All agents report to a human supervisor
4. The largest model supervises smaller ones

Correct answer: A central agent routes requests to specialised sub-agents

In the Supervisor pattern, a central supervisor agent receives requests and routes them to specialised sub-agents based on the task type, then synthesises their responses.