Topic

AI Security

Prompt injection, leakage, unsafe tool access, permissions, and production failure modes.

14 stories (7 articles · 7 videos)

Start here

A few good first pieces before you browse the full feed.

Prompt injection and LLM security: threat models and defense-in-depth

Prompt injection is a permanent LLM security class, not a prompt-writing mistake. A production guide to threat models, data boundaries, tool permissions, regression tests, monitoring, and incident response.

Threat-model an LLM workflow and add concrete controls for untrusted content, retrieval, tool calls, authorization, monitoring, and incident response.

Company knowledge RAG: permissions, leakage, and source boundaries

A company knowledge assistant is only safe if retrieval respects permissions. How to design RAG source boundaries, ACL filtering, document ownership, logging, stale-source handling, and refusal behavior.

Design a company knowledge RAG with permission-aware retrieval, source ownership, leakage controls, and refusal behavior.

Secure document ingestion for RAG: PDFs, OCR, metadata, and retention

RAG quality starts before retrieval. A secure ingestion guide for PDFs, OCR, metadata, permissions, source freshness, deletion, malware risk, and operational ownership.

Design a secure document-ingestion pipeline for RAG with permission metadata, OCR quality checks, source freshness, retention rules, deletion behavior, and ingestion tests.

More in this topic

Unlock Better RAG & AI Agents with Docling

IBM Technology. Explains the ingestion side of RAG and agents: preparing PDFs and other files so document structure, tables and layout survive into downstream retrieval. That supports the article's warning that RAG quality and safety begin before embedding, especially when parsing complex business documents.

Permissions & Access Control for RAG - a Deep Dive Tutorial

Paragon. Walks through the production RAG permission problem and compares tool-calling, namespaces, ACL tables and relationship-based permissions. That directly supports the article's core rule: retrieval must only return sources the current user is allowed to see, and source-system permissions cannot be treated as an afterthought.

How to Build Reliable AI Agents (Context + Evals Explained) | Tobias Leong, Axium

Arize AI. Explains why production agents fail when the system lacks the right context, evaluation data, tracing and domain expertise. It maps well to the article's failure-mode register because it makes reliability an engineering loop: separate retrieval from reasoning, define expected outcomes, evaluate tool calls, and trace failures before changing models.

Production AI failure modes: what breaks after the demo

AI systems usually fail in predictable ways: hallucination, stale context, sycophancy, prompt injection, unsafe tool use, schema drift, and weak fallbacks. A production failure-mode register for teams shipping real workflows.

Build a production AI failure-mode register with controls for hallucination, stale context, prompt injection, unsafe tool use, and weak fallbacks.

Private AI deployment patterns: local, VPC, self-hosted, and hybrid

Private AI is not one architecture. A practical comparison of local models, enterprise SaaS, VPC deployments, self-hosted inference, and hybrid patterns for SMEs that care about privacy and control.

Choose a private AI deployment pattern based on data sensitivity, capability needs, cost, latency, and operational capacity.

Connecting AI to your email, calendar, and CRM safely

Connecting AI to your real tools — email, calendar, CRM — is the productivity unlock and the risk. A practical guide to the integrations that work in 2026, the patterns that are safe, and the lines you should not cross.

Connect AI to email, calendars, and CRMs with least privilege, approval gates, and audit trails.

Privacy and data hygiene when using AI at work

A practical guide to using AI at work without accidentally exposing customer data, breaching your company's policy, or violating GDPR. The lines, the tools, and the habits to build.

Apply practical workplace rules for sensitive data, tool choice, retention, and review before using AI.

Defending LLM - Prompt Injection

LiveOverflow. Walks through the actual defence-in-depth playbook — taint analysis on LLM output, restricting expected output shapes, user isolation, few-shot scaffolds, fine-tuning, temperature 0 for determinism, redundancy for critical paths. It matches the article's defence-stack section almost item for item.

Attacking LLM - Prompt Injection

LiveOverflow. Frames prompt injection as a classic injection attack against systems that mix instructions and untrusted data — with a concrete content-moderation example where an attacker frames an innocent user. The mental shift from "the model is the target" to "the application is the target" is exactly the move the article opens with.

OWASP's Top 10 Ways to Attack LLMs: AI Vulnerabilities Exposed

IBM Technology. Zooms out from prompt injection to the wider OWASP Top 10 for LLMs — insecure output handling, sensitive information disclosure, excessive agency — which is exactly the failure-mode catalogue you want in mind before you grant Gmail or HubSpot scopes to anything.

What Is a Prompt Injection Attack?

IBM Technology. Jeff Crume's "buy an SUV for $1" example is the cleanest 10-minute explanation of why direct and indirect prompt injection are different problems, and why filtering can't fully solve either. It pairs directly with the article's argument that you need least-privilege scopes, a dedicated agent account, and a human in the loop on anything irreversible — not a cleverer system prompt.