Threat-modeling frameworks, model & gateway controls, access controls, data security, monitoring & auditing, and the full catalog of AI-specific attacks with their compensating controls.
Domain 2 is the heart of SecAI+ at 40% of the exam. It covers how to design and enforce controls that protect AI models, data, agents, and integrations — and the full range of AI-specific attacks you're expected to recognize and counter.
Protect AI systems, data, and models using technical safeguards — model controls, gateway controls, and access controls.
Apply best practices across on-premises, cloud, and hybrid AI infrastructure, with monitoring and auditing built in.
Defend against attacks targeting AI models, data pipelines, and inference layers using compensating controls.
Click each card to expand the explanation and study tip.
Apply least privilege across four surfaces:
Agents need particularly tight, scoped permissions — an over-permissioned agent is the root cause of "excessive agency" (see card 6).
This is the longest list on the exam — study attacks and their compensating controls as pairs.
| Attack | What It Does | Compensating Control(s) |
|---|---|---|
| Prompt injection | Malicious instructions embedded in input override system intent | Prompt firewalls, templates, guardrails |
| Model / data poisoning | Corrupts training data or model weights | Data integrity checks, provenance tracking, access controls |
| Jailbreaking | Bypasses a model's safety guardrails | Guardrail testing, layered controls |
| Input manipulation | Crafted inputs cause unintended behavior | Input validation, rate limiting |
| Bias introduction | Deliberately skews training data or outputs | Data auditing, fairness testing |
| Guardrail circumvention | Finds gaps in safety rules | Continuous guardrail validation |
| Integration abuse | Exploits how AI connects to other systems/plug-ins | Least privilege, scoped agent permissions |
| Model inversion / theft | Extracts training data or parameters via queries | Rate limiting, output filtering, access controls |
| Supply chain / transfer learning attack | Compromised pre-trained models or dependencies | Provenance verification, vetted sources |
| Model skewing | Gradually shifts model behavior via crafted inputs over time | Monitoring/auditing for drift |
| Output integrity attack | Tampers with model outputs in transit | Encryption, integrity checks |
| Membership inference | Determines whether specific data was in the training set | Differential privacy, access controls |
| Insecure output handling | Blindly trusting/executing model output (e.g., as code) | Output validation, sandboxing |
| Model denial of service (DoS) | Overwhelms a model with costly queries | Rate/token limits, quotas |
| Sensitive data disclosure | Model reveals confidential training data or PII | Data minimization, redaction, guardrails |
| Insecure plug-ins | Vulnerable third-party extensions | Least privilege, vetting, sandboxing |
| Excessive agency / overreliance | AI agent given too much autonomy or trust | Least privilege, human-in-the-loop oversight |
Click a card to flip it. There are 15 terms covering Domain 2's controls and attacks.
Question 1 of 10 · Score: 0
