PSAA flashcards

38 cards. Read the term and the definition here. The deck at the top of this page is optional.

Prompt injection

Untrusted text that steers a model's behavior. Direct injection comes from user input; indirect injection arrives through retrieved documents, web pages, or tool output. All model input must be treated as attacker-controlled.

Excessive agency

A design flaw where a model holds credentials or permissions broader than its task requires, letting manipulated input trigger destructive or unauthorized actions. Fix by binding tool credentials to the requesting user.

Training-data poisoning

An attacker corrupts the data used to train or fine-tune a model, embedding biased or malicious behavior at the source. Mitigate with data provenance tracking and integrity checks before training.

Retrieval poisoning

Malicious content inserted into a knowledge index or document store so that it surfaces in retrieval-augmented responses. It affects behavior at query time without touching the model weights.

Model output as untrusted

Model responses are data, not commands. Passing them directly into shells, SQL, or downstream APIs without validation creates injection paths equivalent to any other untrusted input.

Zero Trust

A security model that grants no implicit trust based on network location. Every request is authenticated and authorized per session using identity, device posture, and contextual signals.

Microsegmentation

Granular east-west access control that isolates workloads from each other inside a network. It limits lateral movement, which flat internal networks otherwise permit after an initial compromise.

TLS inspection

Decrypting and re-signing traffic to inspect payloads. It requires a trusted re-signing CA on endpoints; otherwise clients reject certificates and inspection silently degrades to passthrough.

IDS versus IPS

An intrusion detection system monitors and alerts on suspicious traffic without blocking it. An intrusion prevention system sits inline and can drop the traffic, so placement determines whether response is possible.

DNS tunneling

Encoding data or command traffic inside DNS queries and responses to bypass port-based controls. Detection relies on logging DNS and flagging anomalous query volume, length, and destination patterns.

Default-deny rule

A policy posture where traffic is blocked unless explicitly allowed. It must be paired with a maintained allow list, since missing entries cause outages rather than security failures.

SASE

Secure Access Service Edge converges network and security services such as SWG, CASB, ZTNA, and FWaaS at a cloud edge near the user, rather than backhauling traffic to a data center.

ZTNA

Zero Trust Network Access brokers per-application access based on identity and device posture. Applications are not exposed to the internet, replacing broad network-level VPN access with narrow authorization.

API-mode CASB

A CASB deployment using provider APIs to inspect SaaS data. It gives deep historical visibility into stored content but is not inline, so it cannot block activity in real time.

Inline proxy CASB

A CASB that sits in the traffic path and can block SaaS actions as they happen. It offers real-time enforcement but sees less historical data than API-based inspection.

Local breakout

Sending branch internet-bound traffic directly to the nearest security point of presence instead of backhauling to a data center. It reduces latency and spreads inspection closer to users.

Single-pass inspection

An architecture where traffic traverses all security engines once in a single pass. It avoids the latency and policy inconsistency introduced by stitching together separate proxy layers.

SWG

Secure Web Gateway filters user web traffic by category, reputation, and risk. In SASE it enforces acceptable-use and malware policy at the cloud edge before traffic reaches the internet.

MTTD and MTTR

Mean time to detect measures event-to-detection latency; mean time to resolve measures detection-to-closure. Tracking them separately exposes whether the bottleneck is visibility or response capacity.

Alert fidelity

The precision of detection alerts, meaning the proportion that represent real incidents. Low fidelity drives analyst fatigue, so tuning focuses on precision rather than raw alert volume.

MITRE ATT&CK

A knowledge base of adversary tactics and techniques based on observed behavior. Defenders use it to map detection coverage and find gaps rather than as a compliance checklist.

Chain of custody

Documented, hash-verified handling of evidence from collection through retention. Gaps in the record can render evidence unusable in investigations or legal proceedings.

Log retention window

The period logs remain queryable. It must cover both detection needs and legal hold, because short retention destroys the ability to investigate intrusions with long dwell time.

Shared responsibility model

The provider secures the cloud itself, including physical and hypervisor layers. The customer secures what runs in the cloud: identity, configuration, data, and access policy.

IAM least privilege

Granting only the actions and resources a principal needs. In practice this means avoiding wildcards, scoping by resource path, and requiring conditions such as MFA where supported.

Public storage exposure

Storage buckets or containers readable or writable by anyone. Block public access at the organization level, then audit individual policies for intentional exceptions.

Policy-as-code guardrails

Organization-level rules that prevent insecure configurations at deployment time. They replace periodic manual review, which detects misconfiguration only after exposure has already occurred.

Workload identity

Short-lived credentials issued to a workload by the platform, replacing static access keys. Role assumption reduces the blast radius of leaked long-lived secrets.

SLO and error budget

An SLO is a reliability target; the error budget is one minus that target. Burn-rate alerting fires on the rate of budget consumption rather than on raw threshold breaches.

Metric cardinality

The number of unique label combinations in a time series. High-cardinality dimensions such as user or request IDs inflate cost and query time, so they belong in logs or traces.

Remediation verification

Confirming a fix with a measurable success signal after action is taken. Without verification, a remediation is only an assumption and the same fault may recur undetected.

Blast-radius limit

A cap on how much an automated remediation can affect, such as one host or one region. It prevents an automated action from turning a small fault into a widespread outage.

Phishing-resistant MFA

Authentication factors such as FIDO2 keys and passkeys that bind to the origin and resist real-time relay. SMS codes are weaker because they can be intercepted or stolen via SIM swap.

Conditional access

Policy that evaluates signals such as user, device compliance, location, and risk at request time. Evaluating only at login misses risk changes that occur later in the session.

OAuth scope over-grant

Issuing delegated tokens with broader scopes than the application needs. Combined with long-lived refresh tokens, this is a common path from a minor app compromise to broad data access.

Just-in-time privilege

Elevating administrative rights only for the duration of a task, with session recording and approval. It removes standing admin access that attackers otherwise harvest and reuse.

Orphaned service account

A machine identity with no owner or active workload. It retains permissions indefinitely, so lifecycle reviews must tie every service account and API key to a responsible team.

Nested group inheritance

Effective permissions accumulated through nested group membership. Access reviews that inspect only direct members miss privileges granted indirectly, so effective access must be calculated.