wayground logo

Free Printable Worksheets

Font size

S
M
L
XL
Worksheets

ECHELON 2K26

Total questions: 25

Worksheet time: 9mins

Name
Class
Date
1.

Which feature most distinguishes modern multimodal AI systems released between 2023–2026?

a)

Ability to run only on mobile CPUs

b)

Ability to jointly process text, images, and other inputs

c)

Ability to compile code automatically

d)

Ability to operate without training data

2.

Which factor most contributes to hallucination reduction in recent LLM architectures?

a)

Higher token limits

b)

Improved beam-search width

c)

Reinforcement learning using verified retrieval sources

d)

More aggressive dropout

3.

What challenge led to the creation of “adaptive context compression” in late-model LLMs?

a)

Slow tokenizer performance

b)

Inability to handle low-resolution images

c)

Lack of GPU availability

d)

Exponentially rising memory cost for long-context attention

4.

Which property becomes most unstable in agentic AI systems executing multi-step tool calls?

a)

Temperature consistency

b)

Action-chain drift

c)

Reward scaling linearity

d)

Tokenization entropy

5.

Continual-retrieval LLMs introduced in 2025–26 rely on which mechanism to avoid catastrophic “knowledge overwrite”?

a)

Dynamic episodic memory buffers

b)

Stochastic pruning

c)

Multi-head redundancy injection

d)

Offline consolidation during decoding

6.

Which mechanism primarily filters packets using predefined allow/deny rules?

a)

Routing table

b)

Firewall

c)

Key-exchange module

d)

Proxy balancer

7.

Which attack overwhelms a target through distributed traffic saturation?

a)

ARP corruption

b)

DNS flattening

c)

Local privilege anomaly

d)

Distributed Denial-of-Service

8.

What principle restricts accounts to only the required authority to perform legitimate tasks?

a)

Least Privilege

b)

Uniform Access

c)

Progressive Trust

d)

Segmented Allocation

9.

Compromising a software dependency such that downstream installations become infected is an example of:

a)

Cache-poison mismatch

b)

Dependency confusion

c)

Socket preemption

d)

Runtime obfuscation

10.

When attackers use built-in OS tools (e.g., shell utilities) to stay stealthy, this is known as:

a)

Reflective tunneling

b)

Silent handshake pivoting

c)

Living-off-the-Land

d)

Credential sublimation

11.

The primary objective of the national ISEA program is to:

a)

Build advanced VR laboratories

b)

Strengthen cybersecurity education & awareness

c)

Develop consumer electronics

d)

Improve mobile broadband penetration

12.

An important ISEA Phase III initiative for universities is the creation of:

a)

AI automation units

b)

Food-processing research cells

c)

Cryptocurrency mining clusters

d)

Cybersecurity labs & cyber ranges

13.

Which demographic, beyond students, is a major target under ISEA Phase III for capacity building?

a)

Government & critical infrastructure workforce

b)

Hospitality professionals

c)

Retail sales staff

d)

Automotive technicians

14.

ISEA Phase III gives significant research attention to which security domain?

a)

Elastic cloud routing

b)

Textile-grade signal modulation

c)

Commodity-priced GPU clusters

d)

Post-quantum cryptography & secure systems

15.

What structural challenge does ISEA Phase III attempt to solve through standardized cyber ranges across institutions?

a)

Lack of uniform, high-fidelity attack simulation environments

b)

Excessive network redundancy across campuses

c)

Overdependence on proprietary GPUs

d)

Low availability of industrial sewing equipment

16.

Which model is considered as AGI?

a)

Gemini 3.0 Pro

b)

Claude Sonnet 4.5

c)

GPT-o4

d)

Human Brain

17.

Which optimization technique adaptively scales the learning rate based on first and second moments of gradients?

a)

Momentum SGD

b)

Adam

c)

Nesterov-only decay

d)

Plain RMSProp

18.

In gradient-boosting frameworks, trees are added sequentially to:

a)

Minimize residual errors from prior models

b)

Decrease depth variance

c)

Reduce global feature entropy

d)

Stabilize minibatch divergence

19.

Which phenomenon occurs when a model trained on one distribution performs poorly because the test distribution subtly shifts, even though labels remain semantically consistent?

a)

Noisy-logit collapse

b)

Covariate shift

c)

Isotropic feature scattering

d)

Loss-surface inversion

20.

Which major tech company recently partnered with the government of United Kingdom to help develop open-source AI tools for improving public services?

a)

Google

b)

Amazon

c)

Meta

d)

IBM

21.

Which mechanisms allow adversarial actors to compromise agentic AI systems performing autonomous tool use?

a)

Prompt-conditioned privilege escalation chains

b)

Multi-step policy hijacking through iterative reward shaping

c)

Memory-constraint overflow produced via long-context injection

d)

Divergence locking by saturating the decoding entropy

22.

Which of the following techniques allow attackers to extract training data from modern LLMs without directly accessing model weights?

a)

Inversion attacks guided by representation embeddings

b)

Regeneration-based data leakage using prompt scaffolds

c)

Beam-search saturation to force deterministic replay

d)

Gradient-alignment extraction using shadow models

23.

Which failure modes arise when multimodal foundation models are targeted with cross-modal adversarial perturbations?

a)

Attention-interval collapse from over-sampling visual tokens

b)

Latent-space skew where internal text/image alignments diverge

c)

Modality-binding desynchronization during fusion-layer aggregation

d)

Sub-token drift caused by vocabulary misalignment

24.

Which defensive strategy demonstrably reduces, but does NOT fully eliminate, training-data poisoning risk in distributed ML pipelines?

a)

Influence-function auditing combined with provenance reconstruction

b)

Batch-normalization freezing to stabilize gradient noise

c)

Sparse regularizers applied to attention projections

d)

Opportunistic teacher–student distillation using noisy priors

25.

Which hybrid AI-driven attack vectors allow malicious actions to blend into legitimate system behavior?

a)

Payload morphing guided by model priors

b)

AI-generated misconfiguration scripts

c)

Semantic perturbation of API calls

d)

Context-aware automated reconnaissance