PSAA key terms

An attack where untrusted input, such as a web page or document, contains instructions that cause a large language model to ignore its original directives and act on the attacker's intent.

A prompt injection delivered through content the model retrieves or processes rather than through the user's own message, such as a poisoned document, email, or web page that carries hidden instructions.

A technique that bypasses a model's safety alignment so it produces content its policy forbids, typically by role-play framing, encoding tricks, or multi-turn persuasion rather than by injecting external content.

Controls placed around a model's input and output to enforce policy, such as content classifiers, schema validation, allow-lists for tools, and blocking of sensitive data patterns before a response is returned.

An architecture where a model is given relevant documents fetched from a search index at query time so answers are grounded in current or private data instead of only pretrained knowledge.

Tampering with training or retrieval data so a model learns wrong associations or surfaces attacker-chosen content, which can later trigger biased answers or malicious actions.

Exposure created by using third-party models, weights, plugins, or datasets whose provenance, licensing, and integrity are not fully verified, allowing hidden backdoors or unsafe behavior.

A capability where a model calls external functions or APIs to act on the world, which expands its reach and makes authorization, input validation, and least privilege essential.

A failure mode where an AI agent is granted permissions, tools, or autonomy beyond what its task requires, so a manipulated or mistaken decision can cause disproportionate harm.

A control requiring a person to review and authorize a high-impact action proposed by an automated system before it executes, limiting damage from errors or manipulation.

A security model that treats no network location as inherently trusted and instead verifies identity and device posture for every request before granting access to a resource.

Dividing a network into small zones with granular allow rules so workloads can reach only the specific peers they need, containing lateral movement after a compromise.

Network flows between internal hosts or workloads inside a data center, as opposed to client-to-server north-south traffic, and a common path for lateral movement.

The techniques an attacker uses after initial access to move from one compromised host to others, escalating privileges and expanding control across the environment.

A sensor that inspects network or host activity against signatures or behavioral baselines and raises alerts on suspicious patterns, either passively or inline to block.

Secure Access Service Edge converges network connectivity such as SD-WAN with cloud-delivered security services like secure web gateway, CASB, and zero trust network access in one policy-managed fabric.

A service that grants users access to individual private applications after identity and device checks, without placing them on the broader corporate network.

A proxy that inspects and filters outbound web traffic, enforcing acceptable-use policy, blocking malicious sites, and applying data controls such as TLS inspection and file scanning.

A cloud access security broker sits between users and cloud services to discover shadow IT, enforce usage policy, and apply data loss prevention and access controls to SaaS applications.

A gateway technique that terminates and re-establishes encrypted sessions so traffic can be examined, requiring certificate trust on endpoints and careful handling of privacy and pinned applications.

Cloud services and applications adopted by employees without IT approval, creating unmanaged data stores and blind spots that discovery tools and policy enforcement aim to surface.

A platform that centralizes logs from many sources, correlates events against detection rules, and alerts analysts, forming the data backbone of a security operations center.

Logic that specifies which event patterns should raise an alert, expressed as signatures, correlation sequences, or behavioral thresholds, and tuned to balance coverage against noise.

A public knowledge base of adversary tactics and techniques observed in real intrusions, used to organize detections, plan threat hunting, and measure defensive coverage gaps.

Proactive, hypothesis-driven searching through telemetry for signs of compromise that existing alerts missed, rather than waiting for a detection to fire.

An observable artifact such as a file hash, domain, or IP address associated with known malicious activity, useful for matching but easily changed by attackers.

An alert raised on benign activity, which consumes analyst time and erodes trust in detection; tuning reduces these while preserving true detections.

A documented, repeatable sequence of steps an analyst or automation follows to triage and respond to a specific alert type, encoding decisions and escalation criteria.

Security orchestration, automation, and response tooling chains together systems and executes playbooks automatically, handling repetitive triage and enrichment tasks to speed response.

The division of security duties between a cloud provider and its customer, where the provider secures the underlying infrastructure and the customer secures data, identities, and configuration.

An insecure setting in a cloud or application service, such as a public storage bucket or overly broad role, that exposes data without requiring an exploit.

A tool that continuously scans cloud accounts against benchmarks and policy, flagging risky configurations and drift and often offering automated remediation.

Static analysis of declarative templates such as Terraform or Kubernetes manifests before deployment, catching insecure resource settings early in the pipeline rather than in production.

The ability to understand a system's internal state from its outputs, combining logs, metrics, and traces so teams can diagnose failures and spot anomalous behavior.

Instrumentation that follows a single request across services using shared trace identifiers, revealing where latency or errors originate in a multi-service architecture.

The average elapsed time from detecting an issue to fully fixing it, a key measure of response effectiveness that complements detection-focused metrics.

The policy defining how long telemetry is stored and at what fidelity, balancing investigation and compliance needs against cost and privacy obligations.

A service that authenticates users and issues assertions or tokens about their identity to relying applications, centralizing credential checks and enabling single sign-on.

Single sign-on lets a user authenticate once with a central identity provider and reach many applications without separate credentials, reducing password sprawl and phishing surface.

Multi-factor authentication requires two or more independent evidence types, such as a password plus a hardware key or push approval, making stolen passwords far less useful.

Credential methods bound to the legitimate service origin, such as passkeys or FIDO hardware keys, which cannot be replayed against a lookalike site by an attacker.

An attack that bombards a user with push notifications until they approve one, or that pairs a stolen password with a single well-timed prompt to gain access.

Policy that grants or blocks access based on context such as user role, device compliance, location, and risk level, evaluated at each sign-in and session.

Granting only the permissions needed for a task and only for as long as needed, limiting the blast radius if an account or workload is compromised.