Security is the protection of systems, data, and operations against unauthorized access, misuse, disruption, or compromise.
Production LLM systems inherit ordinary application-security obligations, but agents add a sharper risk: model-driven execution can expose secrets or trigger harmful actions. section 6.3.6
Prioritize two failure cases:
An attacker persuades an agent to execute malicious code through tools, plugins, browsing, or other integrations. section 6.3.6
An attacker gains access to proprietary data, passwords, secrets, or other sensitive information available to the model. section 6.3.6
Treat prompt injection as a primary threat: malicious instructions can steer a model toward unauthorized actions or disclosure. section 6.3.6
Email agent
An assistant allowed to draft and send email may receive a prompt injection that instructs it to write malware and send it to every contact. Isolation and containment limit the blast radius. section 6.3.6
Apply layered controls:
Sandbox agents according to their use case, especially when they can execute code or act through external tools. section 6.3.6
Retain standard controls such as HTTPS, authentication, authorization, monitoring, logging, network security, and firewalls. section 6.3.6
Use context-aware filtering, including keyword search or a second LLM, as a defense against malicious prompts. section 6.3.6
How LLMs in Production covers it
How deep it goes
Explains ordinary controls, two major failure cases, prompt injection, sandboxing, and filtering without detailing implementation procedures.
Examples
Uses an email-writing agent, tool integrations, internet browsing, malicious prompts, malware, and proprietary secrets.
What it assumes
Assumes familiarity with regular application security while reinforcing its relevance to LLM services.
What you can do after
Operate/configure LLM agents with sandboxing, containment, standard application controls, and prompt-filtering defenses.
How it connects
Connects security with LLM agents, prompt injection, sandboxing, authentication, authorization, logging, and context-aware filtering.
AI for Everyday ITCybersecurity is the practice of protecting digital systems, networks, applications, and data from threats and unauthorized activity.
Build AI-Enhanced Web AppsProtection of generative AI capabilities, application resources, and their outputs against unauthorized access, malicious misuse, and harmful exploitation.
The Art of AI Product DevelopmentCybersecurity is the practice of protecting computer systems, networks, software, and data from unauthorized access, disruption, alteration, or attack.
Effective Data AnalysisData protection is the safeguarding of personal information through controls on its collection, access, processing, storage, sharing, retention, and deletion.
Introduction to Generative AI, Second EditionSecurity is the protection of an AI system, its data, and its interactions from unauthorized access, manipulation, disruption, or attack.
Think Like a CTOAn operational setting in which systems, data, and activities are protected against unauthorized access, misuse, disruption, or compromise.
Machine Learning System DesignSecurity is the protection of a system, its models, data, and interfaces from unauthorized access, misuse, or compromise.
Spring Security in Action, Second EditionSoftware security is the protection of software applications and their data against unauthorized access, misuse, disruption, and other security threats.
Data-Driven CybersecurityInformation security is the protection of information and information systems from unauthorized access, use, disclosure, disruption, modification, or destruction.
Azure SecurityCybersecurity is the practice of protecting digital systems, data, identities, applications, and networks from unauthorized access, disruption, or misuse.
Microsoft Azure in ActionSecurity is the protection of systems, data, applications, and infrastructure against unauthorized access, misuse, disruption, or damage.
Machine Learning Platform EngineeringSecurity is the protection of software, models, data, and infrastructure from unauthorized access, misuse, compromise, or attack.
encryptionuses · Encryption is the cryptographic transformation of readable data into ciphertext to conceal its contents, reversible through decryption with appropriate information.
authorizationuses · Authorization is the determination and enforcement of rules governing which subjects may access resources or perform operations.
authenticationhas part · Authentication is the process of establishing or verifying the identity of a user, service, device, or other communicating entity.
threat modelused by · A threat model is a structured analysis or representation of a system’s relevant threats, adversaries, assets, attack paths, assumptions, and possible mitigations.
AI agentused by · An AI agent is a computational system that uses artificial intelligence to interpret information, select actions, and pursue goals through interaction with an environment.
CIA triadmeasured by · The CIA triad is an information-security model comprising confidentiality, integrity, and availability as core protection objectives.
This explainer was generated by Manning's knowledge graph from the text of LLMs in Production. The section links point to the parts of the book it draws on.
