A formal decision to leave a known security finding unremediated, recorded through instruments such as a risk register, SLA tiers, a "won't fix" state, or an exception and waiver workflow. Accepted risk is a bet on the probability of exploitation, not a claim that the finding is harmless.

The use of malicious inputs or techniques to deceive, manipulate, or exploit AI models. In cybersecurity, adversarial AI can bypass defenses, cause misclassification, or degrade system performance.

The plugins that give AI agents real execution power. Agent Skills are reusable packages of logic, context, and instructions that guide agents through complex workflows. They turn AI from talking to doing.

An AI system capable of autonomous decision-making and taking actions to achieve defined goals without constant human intervention. In cybersecurity, agentic AI can automate complex defense or attack strategies.

The practice of using a team of AI security agents to run an organization's entire application security program in the era of agentic development. It covers the full loop — understanding the application, modeling its threats, finding the vulnerabilities that matter, deciding what is worth fixing, generating and validating fixes, and proving those fixes hold — applied continuously to both new code and the existing backlog.

A software development paradigm where AI agents autonomously plan, execute, and complete development tasks using tools, data, and workflows with minimal human intervention.

A development environment where AI agents do most of the authoring: planning changes, writing code, running tools, and opening pull requests. Security controls that assume a human author reviewing each line do not map onto an ADE.

Multi-step, goal-oriented processes executed by AI agents that orchestrate tools and actions across the SDLC and emerging AI builder workflows.

A software entity powered by artificial intelligence that performs tasks on behalf of a user or system. AI agents in security contexts can monitor threats, respond to incidents, or execute automated workflows.

The ecosystem of tools, integrations, components, and inputs that AI agents rely on to perform tasks. This includes MCP servers, APIs, external services, skills, and other dependencies that influence how agents behave and what they produce.

The risk that is introduced when AI agents consume untrusted or insecure tools, integrations, or inputs. This can lead to data exposure, unintended actions, or the introduction of vulnerabilities during software development.

Cyberattacks enhanced or automated by AI to increase speed, scale, and sophistication, including adaptive phishing, automated exploitation, and evasion.

In AI, bias refers to systemic errors that produce unfair outcomes. Fairness is the mitigation of these biases to ensure equitable and ethical AI performance.

A policy framework outlining principles to protect individuals from harmful AI use. It addresses privacy, discrimination, transparency, and accountability in AI systems.

The elapsed time between an attacker's initial access to a system and the point at which they begin actively exploiting it. As attackers increasingly use AI to automate reconnaissance and exploitation, breakout times have compressed sharply, shrinking the window defenders have to detect and respond.

The protection of AI workloads, models, and data hosted in cloud environments. It combines cloud security practices with AI-specific safeguards against data leakage, model theft, and adversarial attacks.

The automated creation of source code using AI models like LLMs. In cybersecurity, it speeds development but also requires security reviews to prevent vulnerabilities.

AI coding assistants (e.g., Cursor, GitHub Copilot, Claude) are a common and widely adopted form of AI agent. They embed agent capabilities directly into developer workflows, enabling code generation, tool usage, and automated task execution within the development environment. More advanced agents extend beyond coding assistants, operating independently across systems, orchestrating workflows, and interacting with multiple tools without direct human prompting.

The adherence of AI systems to relevant laws, regulations, and industry standards. In cybersecurity, this includes data protection, auditability, and governance frameworks for responsible AI use.

The ability to interpret and understand how an AI model makes decisions. Essential for trust, compliance, and debugging in security systems.

The principle of ensuring that AI systems produce unbiased and equitable outcomes across all user groups. In cybersecurity, AI fairness reduces the risk of discriminatory threat assessments or access controls, supporting compliance and ethical AI deployment.

The process of using a trained AI model to generate predictions, classifications, or responses based on new input data. In cybersecurity, inference powers real-time threat detection, anomaly identification, and automated decision-making without retraining the model.

Unauthorized access, manipulation, or disruption of AI systems, models, or data pipelines. Can lead to compromised outputs, service outages, or exposure of sensitive information.

A method to bypass an AI system’s safety constraints, enabling it to produce restricted or harmful outputs. A critical concern for AI security testing.

The algorithmic structure trained on data to perform tasks such as detection, classification, or generation. Security experts evaluate models for robustness and resistance to attacks.

The use of AI systems to autonomously perform reconnaissance, hypothesis generation, and exploit validation against an application, mirroring the reasoning process of a human penetration tester, rather than pattern-matching against a fixed catalog of known vulnerability signatures.

A proactive security exercise that simulates adversarial attacks against AI systems to identify vulnerabilities and improve resilience.

The process of identifying, evaluating, and mitigating risks related to AI systems. In cybersecurity, it covers model robustness, supply chain vulnerabilities, and compliance risks.

Static application security testing that pairs pattern and dataflow analysis with large language model reasoning, so it can evaluate an application's intent and logic in addition to matching known-bad code patterns.

The integration of AI into security operations to automate threat detection, incident response, and SOC workflows. Enhances speed and accuracy in cybersecurity defense.

Predefined safety and compliance boundaries that prevent AI from generating unsafe or unauthorized outputs. Essential in enterprise AI deployments.

The use of AI tools to proactively search for hidden cyber threats across networks and systems, often detecting patterns missed by human analysts.The use of AI tools to proactively search for hidden cyber threats across networks and systems, often detecting patterns missed by human analysts.

The practice of making AI system operations, decision-making processes, and limitations understandable to stakeholders. In cybersecurity, transparency supports trust, auditing, and regulatory compliance.

Short for Trust, Risk, and Security Management in AI. A governance approach to ensure AI systems are safe, reliable, and compliant.

The confidence stakeholders have in an AI system’s reliability, fairness, and security. Built through transparency, testing, and responsible use.

AI Bill of Materials. A detailed inventory of AI components, datasets, and dependencies. Helps track provenance and manage AI supply chain risks.

The continuous assessment and improvement of an organization’s AI-related security posture, including model governance and threat monitoring.

A living model of an application — its architecture, data flows, data classifications, trust boundaries, and what is actually running in production — that AI security agents read before they act. The graph is built once and updated as the code changes, so every agent reasons from the same understanding of the application rather than blindly scanning.

The technique of combining multiple lower-severity vulnerabilities or logic flaws into a single high-impact exploit path — for example, pairing an authorization gap with a privilege-escalation flaw to reach a critical outcome neither issue alone would allow.

The generation, validation, and delivery of security fixes by AI agents without a human performing each step. Autonomous remediation covers the full loop of scanning, fixing, validating, and opening a pull request a reviewer can merge, with human approval retained wherever an organization wants it.

Amazon Web Services’ tools and frameworks for securing AI workloads, models, and data. Includes threat detection, compliance controls, and AI service protection.

A vulnerability where an application fails to verify that a user is authorized to access a specific object or record, allowing attackers to view or modify data belonging to other users by manipulating an identifier (such as a URL parameter or API request ID). It's the top-ranked risk on the OWASP API Security Top 10.

An assessment of whether applying a fix, typically a dependency upgrade, will break an application's build or behavior. Breakability is what determines whether a correct fix is also a safe one to merge.

A class of vulnerability where an application fails to enforce what an authenticated user may do or see, allowing access to data or actions outside their authorization. It ranks first in the OWASP Top 10 and includes BOLA and IDOR.

A flaw in how an application's intended workflow can be abused, such as accessing another user's data or bypassing an approval step, rather than a flaw in the code itself. Because these vulnerabilities depend on what an application is supposed to do, they have no fixed signature and are difficult for automated scanners to detect without contextual reasoning.

A technique where an attacker or user alters an AI system’s stored conversation history to influence future responses or behaviors. It can be exploited to bypass safeguards, inject malicious instructions, or distort an AI agent’s decision-making

An AI-powered conversational interface that interacts with users via text or voice. In cybersecurity, chatbots can automate support and incident triage.

Non-traditional software builders who leverage AI tools and agents to create or modify applications without formal software engineering expertise. Their rise expands who can build software, increasing both development velocity and security risk.

Discrete instructions or actions that an agent can execute within a system or tool. Commands represent the actual operations performed by the agent (e.g., running a script, querying a service), often composed into multi-step workflows, and can interact with internal or external systems.

Integrated AI architectures combining multiple models or agents to perform complex tasks. In cybersecurity, they enable multi-layered defense strategies.

The process of designing and structuring the context given to AI systems to optimize their outputs. Crucial for accurate cybersecurity automation.

A class of vulnerability that can only be identified by understanding an application's intended behavior — such as authorization logic or multi-tenant data boundaries — as opposed to vulnerabilities with a fixed technical signature that automated scanners can pattern-match directly.

A security testing approach that continuously evaluates applications for exploitable risk — combining automated testing methods like dynamic application security testing (DAST) with deeper, AI-driven penetration testing invoked when warranted — rather than relying on point-in-time assessments run once or twice a year.

Unauthorized use of a system’s resources to mine cryptocurrency, often through malware or browser exploits. AI can both detect and be exploited for such attacks.

Malicious or unethical use of AI for cybercrime, espionage, or disinformation campaigns. Increasingly relevant in threat intelligence.

Unauthorized transfer of data from a system, often by malicious insiders or external attackers. AI can detect anomalies indicating exfiltration attempts.

The process of discovering patterns and insights from large datasets. In cybersecurity, data mining can reveal hidden threats or user behavior anomalies.

A cyberattack strategy that undermines AI model performance by deliberately corrupting datasets through injecting malicious data, mislabeling, or other means.

The safeguarding of data used in AI systems across its lifecycle, including collection, storage, training, and inference. Protects against breaches, leaks, and data poisoning attacks.

A subset of machine learning using neural networks with many layers to model complex data patterns. Powers many AI-driven cybersecurity tools.

Security analysis that returns the same result for the same input on every run, including SAST, SCA, and secrets scanning. In an agentic program, deterministic engines take on two additional roles: the callable tools AI agents invoke, and the independent verifier of what those agents find and fix.

The process of continuously monitoring and improving the security posture of sensitive data across systems and AI workloads.

AI processing that is performed locally on devices rather than in the cloud, reducing latency and enhancing data privacy in cybersecurity applications.

A data integration process that collects, cleans, and moves data into storage or analytics systems. Vital for preparing AI security datasets.

Short for evaluations, these are systematic tests to measure AI system performance, safety, and reliability under varied scenarios.

The ability to interpret and understand how an AI model makes decisions. Essential for trust, compliance, and debugging in security systems.

An assessment of whether a vulnerability can be triggered by an attacker in a specific deployment, accounting for the code path, the configuration, and available exploit techniques.

The process of adapting a pre-trained AI model to a specific task or domain. In security, fine-tuning improves detection accuracy for targeted threats.

The use of a frontier large language model to find vulnerabilities that pattern-based scanners have no signature to match, particularly business logic and authorization flaws. Frontier detection runs alongside deterministic scanning rather than replacing it.

AI models that create new content, such as text, code, or images. Cybersecurity uses include threat simulation and automated phishing generation.

A framework for managing an organization’s governance, risk management, and regulatory compliance, now extended to AI systems.

When an AI model generates false or misleading outputs that appear plausible. In security, hallucinations can lead to misinformation or faulty automation.

A credential such as an API key, token, password, or certificate written directly into source code, configuration, or commit history instead of a managed secret store. Detection combines entropy analysis, machine learning, and pattern matching across the development lifecycle.

Predefined triggers that execute specific logic at certain points in an agent’s workflow. Hooks run automatically before, during, or after an action, often operating transparently within agent execution. Can be used to inject validation, transformation, or enforcement logic.

This term refers to human involvement in machine learning processes and AI workflows. Human experts provide insights, expertise, and judgment to

Combines multiple AI approaches—such as symbolic reasoning and machine learning—for better accuracy and adaptability in cybersecurity.

A vulnerability where an application exposes a reference to an internal object, such as a record ID, filename, or key, without verifying the requester's authorization. An attacker substitutes another value to reach data belonging to someone else.

A centralized hub for developers to access tools, documentation, and services. In security, it ensures controlled access to AI development resources.

The practice of validating a security fix with a system other than the one that generated it. A model that grades its own output cannot provide an independent assessment, so validation is delegated to a deterministic engine or a separate validation system.

An AI framework for building applications powered by large language models with memory, chaining, and context control capabilities.

A machine learning model trained on massive datasets to understand and generate human-like text. Used in code generation, analysis, and chatbots.

The unauthorized hijacking or exploitation of large language models, often to inject malicious prompts, steal intellectual property, or manipulate outputs for cyberattacks.

The operational discipline of deploying, monitoring, and maintaining large language models in production environments.

A package published to a public registry with the intent to compromise whoever installs it, through credential theft, data exfiltration, or code execution in install scripts. Install-time defense blocks these before they reach a developer machine or build environment.

The Model Context Protocol is an open standard that allows developers to create secure, bidirectional connections between AI-powered tools and their data sources.

A security fix a developer accepts and merges, as distinct from one that is merely generated. A fix is mergeable when it resolves the underlying issue, does not break the build, does not cascade unwanted dependency changes, and carries enough evidence for a reviewer to approve it.

The MITRE Adversarial Threat Landscape for Artificial-Intelligence Systems, a knowledge base of adversary tactics, techniques, and case studies for AI systems. Guides security teams in defending AI assets.

A subset of AI that uses algorithms to learn from data and make predictions or decisions without explicit programming.

Practices for managing the lifecycle of machine learning models, from training to deployment, ensuring reliability and scalability.

The integration of machine learning, security, and operations practices to safeguard AI systems throughout their lifecycle.

The unauthorized extraction or replication of a proprietary AI model, often for competitive or malicious purposes.

A computational model inspired by the human brain, used in deep learning to process complex data relationships.

AI techniques that allow computers to understand, interpret, and generate human language. Used in chatbots, threat detection, and intelligence gathering.

AI systems whose outputs can vary for the same input due to probabilistic processes, requiring special handling in security contexts.

When an AI system fabricates the existence of software packages or components, leading to potential supply chain risks.

The integration of AI capabilities across all aspects of an organization’s operations, tools, and infrastructure. In cybersecurity, pervasive AI enables continuous, automated defense at scale.

AI systems designed to forecast outcomes or behaviors based on historical data patterns. Used in threat prediction and risk assessment.

The practice of crafting effective prompts to guide AI model outputs, optimizing performance for security use cases.

A malicious attempt to manipulate an AI system’s behavior by embedding harmful instructions into input prompts.

An AI approach that combines language generation with external knowledge retrieval for more accurate and context-rich answers.

An assessment of whether a vulnerable function in a dependency is actually called by an application's code paths. Reachability separates findings that are theoretically present from those that an application can execute.

A machine learning approach where an agent learns through trial and error by receiving rewards or penalties, and its goal is to learn a strategy to maximize its cumulative reward over time without explicit instructions or labeled data

The development and deployment of AI systems that are ethical, fair, transparent, and secure. In cybersecurity, responsible AI practices reduce bias, ensure compliance, and minimize misuse risks.

Defined constraints or policies that can be applied across agent workflows and that guide how an agent should behave when performing tasks. Rules can specify allowed or disallowed behaviors, enforce security, compliance, or operational requirements.

The principle of ensuring that agent-generated outputs are secure as part of the agentic development process, through embedded security intelligence and guardrails within agent workflows, reducing reliance on human intervention for validating newly generated code.

A moment in the development lifecycle where security analysis runs. Three now operate at different scopes: real time inside the coding agent's loop, deterministic scans on every change in CI, and deep contextual analysis across the whole codebase on a cadence. Each catches a different class of issue.

The unapproved use of AI tools and models within an organization, creating compliance and security risks.

The malicious registration of packages with names similar to legitimate ones to trick developers into installing them.

Specialized agents that are delegated specific tasks as part of a larger workflow. Subagents operate with a defined scope or function, often collaborating with other agents to complete complex processes, and inheriting access to tools, data, and permissions.

A structured analysis of an application's architecture, data flows, and trust boundaries to identify where it can be attacked and which weaknesses matter most. Agentic threat modeling derives and updates this model continuously from the application context graph instead of producing it in a one-off workshop.

A type of data or model attack where malicious inputs are inserted into AI tools or datasets to manipulate behavior or outcomes. It compromises the reliability of AI systems over time.

The dataset used to teach an AI model how to perform a task. In security, its quality directly impacts detection accuracy.

A new approach to software development where developers create code through AI-powered platforms simply by describing desired outcomes in natural language, without requiring manual coding or prior programming knowledge.

The accumulated set of known, unremediated security findings that an organization carries. When remediation trails are introduced, the backlog serves as a standing inventory of exploitable paths that automated attackers can enumerate, rather than as deferred maintenance work.

The accumulated set of known, unremediated security findings that an organization carries. When remediation trails are introduced, the backlog serves as a standing inventory of exploitable paths that automated attackers can enumerate, rather than as deferred maintenance work.