PAPA study plan

Define the core AI security threat categories (prompt injection, data poisoning, model extraction, adversarial input, insecure output handling) and write a one-line definition plus a concrete example for each

Map the AI attack surface across the model lifecycle (data collection, training, deployment, inference) and label where each threat category applies

Explain the difference between direct and indirect prompt injection and identify which one crosses a trust boundary in a RAG pipeline

List the OWASP Top 10 for LLM Applications categories from memory and match each to a mitigation pattern

Compare guardrail controls (input filtering, output validation, system prompt hardening, allowlisting) and state the failure mode each control does not cover

Describe how model supply chain risk arises from pretrained weights, third-party plugins, and fine-tuning datasets, and outline verification steps for each

Explain data poisoning detection approaches and distinguish them from normal model drift monitoring

Draft least-privilege rules for an AI agent that can call tools, including scoping, human approval gates, and rate limits

Identify privacy risks in training and inference (memorization, PII leakage in prompts, logging exposure) and propose two controls for each

Diagram the OSI and TCP/IP layers and place at least 15 common attacks (ARP spoofing, SYN flood, DNS poisoning, TLS downgrade, VLAN hopping) at the layer where they occur

Explain how TLS handshakes establish trust and identify where certificate pinning, HSTS, and mutual TLS change the threat model

Compare stateful vs stateless firewall behavior and describe how a next-generation firewall differs in application inspection

Describe network segmentation and zero trust architecture principles and contrast them with perimeter-based defense

Explain IDS vs IPS placement, signature vs anomaly detection, and the false positive tradeoffs of each

Explain how DNS resolution, DNSSEC, and DNS-over-HTTPS work and where cache poisoning and tunneling attacks fit

Describe common VPN and remote access weaknesses and the controls that mitigate them

Analyze a packet capture scenario by reasoning about TCP flags, retransmissions, and abnormal port behavior without relying on tooling

Compare wireless security protocols (WPA2, WPA3) and identify the attacks each resists or fails to resist

Summarize how BGP hijacking and route manipulation can redirect traffic and what detection looks like

Describe the SOC tier model and the escalation criteria that move an alert from triage to incident response

Walk through the incident response lifecycle phases and state the key decision and artifact produced in each

Explain log source categories (endpoint, network, identity, cloud audit) and what each is best at detecting

Practice writing detection logic in a SIEM query language for three scenarios: brute force, impossible travel, and data exfiltration volume spike

Distinguish indicators of compromise from indicators of attack and explain why the latter is harder to act on

Explain the MITRE ATT&CK tactic and technique structure and map three real attack scenarios to technique IDs

Describe threat hunting hypotheses and how they differ from alert-driven monitoring

Outline forensic evidence handling: order of volatility, chain of custody, and memory vs disk acquisition tradeoffs

Explain how SOAR playbooks reduce mean time to respond and where automation creates risk

Summarize post-incident review outputs and how lessons learned feed back into detection content

Explain the shared responsibility model and locate the boundary for IaaS, PaaS, and SaaS across compute, network, identity, and data

Describe identity and access management risks in cloud: overprivileged roles, long-lived keys, confused deputy, and cross-account trust

Explain how storage misconfiguration leads to public exposure and the controls that prevent and detect it

Compare network controls in cloud (security groups, NACLs, private endpoints, egress filtering) and state when each applies

Describe logging and monitoring coverage in cloud environments and identify common blind spots such as control-plane vs data-plane events

Re-derive the weakest two concepts from each domain using your own notes and confirm understanding without reference material

Integrate the domains by tracing one end-to-end attack that starts with a network foothold, escalates through cloud identity, abuses an AI service, and is detected in the SOC

Complete two timed mixed-domain practice sets and record per-domain accuracy

Review every missed item and write the corrected reasoning for each in one sentence