AI Security

The AI Security desk

Deep dives on the OWASP Top 10 for LLM applications, and a running record of the real-world AI security incidents of the last four years — what happened, why, and the lesson every security team should take from it.

OWASP Top 10 for LLM Applications

The canonical risk framework for LLM apps, one clear explainer per category.

Attack Surface

Shadow AI: The Invisible Attack Surface

Unsanctioned LLMs, agents, and MCP servers are already in your fleet — here's why discovery has to come first.

Aug 13, 20268 min read
Read →
LLM01

Prompt Injection, Explained

Direct vs. indirect injection, real-world examples, and why OWASP ranks it as the number-one LLM risk.

Aug 8, 20269 min read
Read →
Regulation

The EU AI Act: What Security Teams Actually Need to Do

The adversarial-testing duty, the evidence you'll be asked for, and the timelines that are already live.

Aug 1, 20268 min read
Read →
MCP

Securing MCP Servers and AI Agents

Over-broad tool scope, typosquatting, and tool poisoning — plus a hardening checklist you can run today.

Jul 25, 20269 min read
Read →
LLM02

Sensitive Information Disclosure in LLM Applications

The second-ranked LLM risk is also the quietest: models that leak the very data they were trusted to handle.

Aug 13, 20268 min read
Read →
RAG

What RAG Really Is — and Why You Need a Vector Database

Retrieval-augmented generation, the vector database it runs on, and the security seams that open up the moment you bolt your data onto a model.

Aug 14, 20269 min read
Read →
Confused Deputy

The Confused Deputy Problem in AI Agents

A decades-old security flaw is having a renaissance — because an AI agent is the most powerful confused deputy we have ever built.

Aug 14, 20268 min read
Read →
MCP

Why MCP, When You Can Just Call Tools Directly

The Model Context Protocol exists to kill the N×M integration problem — and in doing so it turns every tool connection into a shared trust surface.

Aug 5, 20268 min read
Read →
Orchestrator

Why an Orchestrator, When You Already Have MCP

MCP gives an agent hands. The orchestrator is the nervous system that decides when to use them — which makes it the component that holds the credentials and the blast radius.

Aug 7, 20268 min read
Read →
Trust Boundary

How an LLM Application Actually Works — Architecture and Trust Boundaries

An LLM application is not just a model — it is a prompt-assembly pipeline with one critical trust boundary most teams draw in the wrong place.

Jul 28, 20269 min read
Read →
Agentic Loop

The Agentic Loop and Where It Breaks

Perceive, plan, act, observe — the loop that makes an agent an agent is also the mechanism that lets one poisoned observation hijack everything that follows.

Aug 9, 20268 min read
Read →
Lethal Trifecta

The Lethal Trifecta: Private Data, Untrusted Content, and an Exfiltration Path

Three agent capabilities that are each harmless alone become a data-theft machine the moment one system has all three at once.

Aug 11, 20269 min read
Read →
Tool Poisoning

Tool Poisoning and Malicious MCP Servers

An MCP tool description is prose the model obeys — which makes the server that writes it a place to hide instructions your agent will follow.

Aug 6, 20268 min read
Read →
Embeddings

Embeddings and the Vector-Database Threat Model

Vectors are not anonymised numbers — they are a lossy copy of your text, stored in a database that usually has none of your usual access controls.

Jul 30, 20269 min read
Read →
Context Window

The Context Window as a Trust Boundary

Everything the model reads — your rules, the user's text, retrieved docs, tool output — lands in one flat buffer with no privilege separation. That is the whole problem.

Aug 12, 20268 min read
Read →
Supply Chain

The AI Supply Chain: Models, Datasets, Adapters and Plugins

Every model you ship is assembled from parts you did not build — scraped data, downloaded weights, community adapters and third-party plugins. Each is an injection point.

Aug 15, 20269 min read
Read →
Agent Auth

Agent Identity and Delegated Authorization

An agent acting for a user needs authority without becoming the user. That is the problem OAuth solved for apps — and the one we keep re-solving badly for agents.

Aug 16, 20268 min read
Read →
Agent Memory

Agent Memory and Persistence Risks

Memory is what makes an agent feel personal — and what lets a single poisoned input follow a user across every future session.

Aug 17, 20268 min read
Read →
Guardrails

Guardrails: What They Can and Cannot Do

Guardrails are useful and oversold in equal measure. Knowing precisely what they can enforce — and what they structurally cannot — is the difference between defence and theatre.

Aug 18, 20268 min read
Read →
Non-Determinism

Non-Determinism as a Security Property

Traditional testing assumes the same input yields the same output. AI breaks that assumption — and with it, the meaning of a passing test.

Aug 19, 20268 min read
Read →
Evals

Evaluations and Red-Teaming as Security Controls

If a passing test is only a sample, you need two disciplines to know an AI system is safe: repeatable evaluations for the known, red-teaming for the unknown.

Aug 20, 20268 min read
Read →
MCP

MCP Security Architecture: Authentication, Scopes and Transport

The Model Context Protocol standardised how agents reach tools and data. Standardising the plumbing also standardised the attack surface — here is where to secure it.

Aug 21, 20269 min read
Read →
LLM03

Supply Chain Risk in the LLM Stack

The model you shipped is something you assembled, not something you built — and every borrowed piece is a link an attacker can pull.

Aug 11, 20268 min read
Read →
LLM04

Data and Model Poisoning

Corrupt the data a model learns from and you corrupt the model — quietly, durably, and in ways that survive every later test you didn't think to run.

Aug 9, 20268 min read
Read →
LLM05

Improper Output Handling: When Model Output Becomes an Exploit

A model's output is untrusted input to whatever comes next. Treat it as trusted and you hand classic injection bugs a brand-new source.

Aug 7, 20268 min read
Read →
LLM06

Excessive Agency: When Your AI Can Do Too Much

Give a probabilistic system tools, permissions, and autonomy, and its worst mistake is bounded only by what you let it reach.

Aug 5, 20268 min read
Read →
LLM07

System Prompt Leakage: Secrets You Should Never Have Put There

The danger isn't that someone reads your system prompt — it's what teams keep hiding inside it, assuming no one ever will.

Aug 3, 20267 min read
Read →
LLM08

Vector and Embedding Weaknesses: The Security of RAG

Retrieval-augmented generation moved your sensitive data into a vector store — and most of the access controls stayed behind.

Aug 1, 20268 min read
Read →
LLM09

Misinformation and Hallucination as a Security Risk

A confidently wrong model isn't just a quality problem — it's a liability, an attack surface, and, in code, a fresh supply-chain hole.

Jul 30, 20268 min read
Read →
LLM10

Unbounded Consumption: Denial of Wallet and Model Theft

Metered, expensive inference turns old-fashioned resource abuse into a bill — and gives attackers a way to steal the model one query at a time.

Jul 28, 20268 min read
Read →
Incident file · 2022–2025

Documented, publicly-reported incidents — reported factually, with the takeaway that matters.

Incident · 2024

Air Canada and the Chatbot That Cost the Airline in Court

A tribunal ruled an airline liable for its chatbot's bad advice — and rejected the idea that the bot was a separate legal entity.

Aug 12, 20266 min read
Read →
Incident · 2023

Samsung, ChatGPT, and Three Leaks in Twenty Days

Weeks after Samsung let engineers use ChatGPT, staff pasted semiconductor source code and meeting notes into it three separate times.

Jul 15, 20267 min read
Read →
Incident · 2023

The Redis Bug That Showed ChatGPT Users Each Other's Chats

For a few hours in March 2023, a caching race condition let ChatGPT users see strangers' conversation titles and, for some, partial payment data.

Jul 18, 20267 min read
Read →
Incident · 2023

Microsoft's AI Team, One SAS Token, and 38TB of Exposure

A single misconfigured Azure token in a public AI repo exposed 38 terabytes of internal data — including workstation backups and secrets — for years.

Jul 21, 20268 min read
Read →
Incident · 2023

PoisonGPT: A Lobotomised Model Hidden in Plain Sight

Researchers surgically edited an open model to lie about specific facts, uploaded it under a look-alike name, and showed it passed standard benchmarks.

Jul 24, 20267 min read
Read →
Incident · 2023

1,600 Leaked Tokens and the Soft Underbelly of the AI Supply Chain

Lasso Security found over 1,600 valid Hugging Face tokens exposed in public code, many with write access to models from Meta, Google, and Microsoft.

Jul 27, 20267 min read
Read →
Incident · 2023

“You Are Sydney”: How Bing Chat Gave Up Its System Prompt

Days after launch, a student coaxed Microsoft's new Bing Chat into reciting the confidential instructions it had been told never to reveal.

Jul 30, 20266 min read
Read →
Incident · 2023

When Italy Pulled the Plug on ChatGPT

In March 2023, Italy's data regulator became the first in the West to block ChatGPT, turning AI privacy from an abstract worry into an operational one.

Aug 2, 20266 min read
Read →
Incident · 2023

The Chatbot That Agreed to Sell a Chevy Tahoe for a Dollar

A dealership bolted ChatGPT onto its website. Within a viral afternoon, users had it agreeing to $1 SUVs and answering questions about rival brands.

Aug 5, 20266 min read
Read →
Incident · 2023

The Invisible Pixel: Stealing Data Through a Chatbot's Images

Researcher Johann Rehberger showed how a rendered Markdown image could quietly smuggle a user's private data out to an attacker's server.

Aug 7, 20267 min read
Read →
Incident · 2023

When ChatGPT Invents a Package Name, Attackers Register It

Vulcan Cyber showed that AI coding assistants confidently recommend software packages that don't exist — and that an attacker can register the name and wait.

Aug 9, 20267 min read
Read →
Incident · 2023

WormGPT and FraudGPT: Crime-as-a-Service Gets a Language Model

In mid-2023, dark-web sellers began advertising ChatGPT clones with the safety filters stripped out, purpose-built for phishing and fraud.

Aug 11, 20267 min read
Read →
Incident · 2023

DAN and the Jailbreak Arms Race

A crowdsourced roleplay prompt called DAN spent 2023 trying to talk ChatGPT out of its own safety rules — and kept evolving as OpenAI patched it.

Aug 12, 20266 min read
Read →
Incident · 2023

100,000 Stolen ChatGPT Logins on the Dark Web

Group-IB found over 100,000 ChatGPT credentials in infostealer logs — not because ChatGPT was breached, but because of what users had typed into it.

Aug 4, 20266 min read
Read →
Incident · 2023

From Prompt Injection to Code Execution: The MathGPT Case

A public AI app that turned math questions into Python was talked into running the attacker's Python instead — leaking its own API key.

Aug 6, 20267 min read
Read →
Incident · 2023

Indirect Prompt Injection: Attacks Hidden in the Page

Researchers showed Bing's chat could be hijacked not by what the user typed, but by invisible text on a web page it happened to read.

Aug 10, 20267 min read
Read →
Incident · 2024

DPD's Chatbot Swore at a Customer and Wrote Poems Against Its Own Company

A UK delivery firm's support bot was talked into cursing and mocking its employer — a lesson in what happens when an update quietly removes guardrails.

Jul 15, 20266 min read
Read →
Incident · 2024

New York City's Chatbot Told Businesses to Break the Law

NYC's official MyCity chatbot advised employers and landlords to do things that are plainly illegal — and stayed online after it was exposed.

Jul 18, 20267 min read
Read →
Incident · 2024

Slack AI and the Prompt Injection That Reached Into Private Channels

PromptArmor showed how a message in a public Slack channel could coax Slack AI into leaking data from a private one via indirect prompt injection.

Jul 21, 20268 min read
Read →
Incident · 2024

Living off Microsoft Copilot: Turning an Assistant Into an Insider

At Black Hat 2024, Zenity's Michael Bargury showed how prompt injection could bend Microsoft 365 Copilot into a phishing and data-extraction tool.

Jul 24, 20268 min read
Read →
Incident · 2024

ShadowRay: When Exposed AI Compute Clusters Became a Cryptomining Farm

Oligo found thousands of internet-exposed Ray clusters being exploited — amid a dispute over whether it's a vulnerability or the framework working as designed.

Jul 27, 20268 min read
Read →
Incident · 2024

Malicious Models and Shared Tenancy: Wiz's AI-as-a-Service Research

Wiz uploaded malicious models to Replicate and SAP AI Core to cross tenant boundaries — showing a model file is executable code, not just data.

Jul 30, 20268 min read
Read →
Incident · 2024

The Ultralytics YOLO Compromise: A Poisoned Build Pipeline on PyPI

Malicious versions of Ultralytics YOLO shipped a cryptominer to PyPI — not by stealing a password, but by poisoning the GitHub Actions build cache.

Aug 2, 20267 min read
Read →
Incident · 2025

DeepSeek's Exposed Database and the Cost of a Missing Password

As DeepSeek's models went viral, Wiz found one of its databases open to the internet with no authentication — plaintext chat logs and secret keys included.

Aug 5, 20267 min read
Read →
Incident · 2024

ChatGPT's macOS App Stored Your Conversations in Plaintext

A developer found OpenAI's ChatGPT Mac app saving every conversation in unencrypted local files, readable by any other app on the machine.

Aug 7, 20266 min read
Read →
Incident · 2024

Vanna.AI: When a Text-to-SQL Assistant Becomes Remote Code Execution

JFrog showed how a prompt to the Vanna.AI library could jump the gap from natural-language question to arbitrary code execution — CVE-2024-5565.

Aug 9, 20267 min read
Read →
Incident · 2024

Microsoft Recall and the Screenshot Database Nobody Asked to Encrypt

Microsoft's Recall feature captured everything on your screen into a local, unencrypted database — until a security backlash forced a redesign.

Aug 11, 20267 min read
Read →
Incident · 2025

EchoLeak: The Zero-Click Flaw in Microsoft 365 Copilot

Aim Security disclosed a zero-click vulnerability in Microsoft 365 Copilot that could exfiltrate a user's data from a single unopened email — CVE-2025-32711.

Aug 12, 20268 min read
Read →
Incident · 2024

Gemini's Image Generator and the Governance Lesson Google Learned in Public

Gemini's image generator produced historically inaccurate depictions of people and Google paused it — a case study in AI governance, not a breach.

Aug 13, 20266 min read
Read →
Incident · 2024

The OpenAI Breach You Did Not Hear About Until a Year Later

A hacker breached OpenAI's internal employee messaging system in early 2023 — a fact the public did not learn until The New York Times reported it in July 2024.

Aug 14, 20267 min read
Read →
Incident · 2025

nullifAI: Malicious Models That Slipped Past Hugging Face's Scanner

ReversingLabs found malicious models on Hugging Face that hid a payload inside a deliberately broken pickle file to evade the platform's security scanner.

Aug 14, 20267 min read
Read →