Font size
WorksheetsECHELON 2K26
Total questions: 25
Worksheet time: 9mins
Which feature most distinguishes modern multimodal AI systems released between 2023–2026?
Ability to run only on mobile CPUs
Ability to jointly process text, images, and other inputs
Ability to compile code automatically
Ability to operate without training data
Which factor most contributes to hallucination reduction in recent LLM architectures?
Higher token limits
Improved beam-search width
Reinforcement learning using verified retrieval sources
More aggressive dropout
What challenge led to the creation of “adaptive context compression” in late-model LLMs?
Slow tokenizer performance
Inability to handle low-resolution images
Lack of GPU availability
Exponentially rising memory cost for long-context attention
Which property becomes most unstable in agentic AI systems executing multi-step tool calls?
Temperature consistency
Action-chain drift
Reward scaling linearity
Tokenization entropy
Continual-retrieval LLMs introduced in 2025–26 rely on which mechanism to avoid catastrophic “knowledge overwrite”?
Dynamic episodic memory buffers
Stochastic pruning
Multi-head redundancy injection
Offline consolidation during decoding
Which mechanism primarily filters packets using predefined allow/deny rules?
Routing table
Firewall
Key-exchange module
Proxy balancer
Which attack overwhelms a target through distributed traffic saturation?
ARP corruption
DNS flattening
Local privilege anomaly
Distributed Denial-of-Service
What principle restricts accounts to only the required authority to perform legitimate tasks?
Least Privilege
Uniform Access
Progressive Trust
Segmented Allocation
Compromising a software dependency such that downstream installations become infected is an example of:
Cache-poison mismatch
Dependency confusion
Socket preemption
Runtime obfuscation
When attackers use built-in OS tools (e.g., shell utilities) to stay stealthy, this is known as:
Reflective tunneling
Silent handshake pivoting
Living-off-the-Land
Credential sublimation
The primary objective of the national ISEA program is to:
Build advanced VR laboratories
Strengthen cybersecurity education & awareness
Develop consumer electronics
Improve mobile broadband penetration
An important ISEA Phase III initiative for universities is the creation of:
AI automation units
Food-processing research cells
Cryptocurrency mining clusters
Cybersecurity labs & cyber ranges
Which demographic, beyond students, is a major target under ISEA Phase III for capacity building?
Government & critical infrastructure workforce
Hospitality professionals
Retail sales staff
Automotive technicians
ISEA Phase III gives significant research attention to which security domain?
Elastic cloud routing
Textile-grade signal modulation
Commodity-priced GPU clusters
Post-quantum cryptography & secure systems
What structural challenge does ISEA Phase III attempt to solve through standardized cyber ranges across institutions?
Lack of uniform, high-fidelity attack simulation environments
Excessive network redundancy across campuses
Overdependence on proprietary GPUs
Low availability of industrial sewing equipment
Which model is considered as AGI?
Gemini 3.0 Pro
Claude Sonnet 4.5
GPT-o4
Human Brain
Which optimization technique adaptively scales the learning rate based on first and second moments of gradients?
Momentum SGD
Adam
Nesterov-only decay
Plain RMSProp
In gradient-boosting frameworks, trees are added sequentially to:
Minimize residual errors from prior models
Decrease depth variance
Reduce global feature entropy
Stabilize minibatch divergence
Which phenomenon occurs when a model trained on one distribution performs poorly because the test distribution subtly shifts, even though labels remain semantically consistent?
Noisy-logit collapse
Covariate shift
Isotropic feature scattering
Loss-surface inversion
Which major tech company recently partnered with the government of United Kingdom to help develop open-source AI tools for improving public services?
Amazon
Meta
IBM
Which mechanisms allow adversarial actors to compromise agentic AI systems performing autonomous tool use?
Prompt-conditioned privilege escalation chains
Multi-step policy hijacking through iterative reward shaping
Memory-constraint overflow produced via long-context injection
Divergence locking by saturating the decoding entropy
Which of the following techniques allow attackers to extract training data from modern LLMs without directly accessing model weights?
Inversion attacks guided by representation embeddings
Regeneration-based data leakage using prompt scaffolds
Beam-search saturation to force deterministic replay
Gradient-alignment extraction using shadow models
Which failure modes arise when multimodal foundation models are targeted with cross-modal adversarial perturbations?
Attention-interval collapse from over-sampling visual tokens
Latent-space skew where internal text/image alignments diverge
Modality-binding desynchronization during fusion-layer aggregation
Sub-token drift caused by vocabulary misalignment
Which defensive strategy demonstrably reduces, but does NOT fully eliminate, training-data poisoning risk in distributed ML pipelines?
Influence-function auditing combined with provenance reconstruction
Batch-normalization freezing to stabilize gradient noise
Sparse regularizers applied to attention projections
Opportunistic teacher–student distillation using noisy priors
Which hybrid AI-driven attack vectors allow malicious actions to blend into legitimate system behavior?
Payload morphing guided by model priors
AI-generated misconfiguration scripts
Semantic perturbation of API calls
Context-aware automated reconnaissance
