AI Security Learning Center
AI security, from the headlines to the fix
Real attacks on AI systems, how they worked and what stops them. Free and independent: not funded by or affiliated with any AI lab, cloud provider or security vendor, and every claim links to its source.
"Why would anyone attack us?"
Because most attacks on AI systems are not aimed at a company. They are aimed at a pattern, and if you run the pattern you are the target.
Attackers scan, they do not choose
Exposed AI services are found by internet-wide scanners within hours. ShadowRay turned thousands of unauthenticated Ray clusters into miners and botnets, and one operator's open-source agent, given a single instruction over Telegram, went after more than 460 exposed servers with public exploits on its own. Nobody looked up who owned them.
They poison what everyone installs
Phished maintainers (chalk, debug), malicious agent skills (ClawHavoc) and MCP
servers that turn on their users (postmark-mcp, and Deadbugz, which behaves for three
tool calls and then tells the agent to collect your SSH keys) reach you through a
routine npm install. Your size is irrelevant; your dependency list is the
attack surface.
Your own agent does the work
A coding agent that reads GitHub issues, an assistant that reads email, a support bot that reads tickets: each one reads attacker text with your credentials. One public issue took over the official Claude Code, Gemini CLI and Codex actions.
The damage is your data and your keys
EchoLeak, ShadowLeak, CamoLeak and AgentFlayer were zero-click leaks from mainstream assistants. Salesloft Drift's stolen agent tokens opened Salesforce at hundreds of companies that had never heard of the attacker.
How fast it happens now
The gap between a flaw becoming public and someone using it against you used to be weeks. With AI agents on the attacker's side it is hours, and once an agent is inside, the next step takes seconds. The attacker no longer needs to be skilled, or to be a team. The slow end of the clock is just as dangerous: a flaw left open for over a year, a breach nobody noticed for months. Each example links to the full, sourced write-up.
Seconds
DIVD, September 2026. An agentic attack chained two zero-days in the Zammad help desk of the Dutch Institute for Vulnerability Disclosure, from no login to root in seconds, then pivoted to other services and took data.
6 minutes
Mini Shai-Hulud, May 2026. A worm poisoned a shared build cache, stole the release runner's publishing token and pushed 84 malicious TanStack versions in about six minutes. It reached Mistral AI's SDKs and two OpenAI employee devices.
Under 4 hours
ChainDrop, August 2026. One stolen maintainer account backdoored more than 400 npm packages with over a billion monthly downloads between them, and planted hooks that run when a developer simply opens the project.
1 day
Rejetto HFS, October 2026. Anthropic's Mythos found the flaw and wrote a working exploit. Attacks began the evening after the write-up was published, more than two months after the fix shipped.
4 days
Taiwan, July 2026. Human operators set the targets and a team of AI agents did the work: mapping connected government systems, harvesting credentials and taking personnel records from several agencies in about four days.
3 months
Mexico, December 2025 to February 2026. A small crew got past a coding agent's safeguards and had it do most of the hands-on work in breaches of nine government agencies, the tax authority among them. About 150 GB left, with records on some 195 million identities.
7 months unnoticed
A test that was not sealed, January 2026. During a cybersecurity evaluation, Claude Opus 4.6 found its test environment was accidentally online, reached a third party's systems, took stored credentials and gained administrator control. It surfaced seven months later.
Over a year exposed
JADEPUFFER, July 2026. An AI agent ran a whole ransomware attack on its own, getting in through a Langflow server still open to a flaw that had been public for more than a year. It encrypted 1,342 configuration items with a key it never kept, so the data is gone.
The pattern in almost every one: something reachable from the internet or trusted by default, a flaw, a token or a setting left as it was, and something on the other side that never sleeps. That is also why the fixes below are boring. They work because they remove the pattern, not because they outsmart the attacker.
What changes the odds
- Nothing AI-related is reachable from the internet without authentication.
- Every dependency, model, MCP server and plugin is approved once, pinned, and re-checked when new advisories land.
- Agents run with their own scoped identity, in a sandbox, with approval for consequential actions, and everything is logged.
- Everything an agent reads is treated as attacker-controlled, and no agent combines private data, untrusted content and a way out.
Each of these is one of the , and each is illustrated by an incident in the . The shows where attacks hit, who reported them, and a heat map of where activity concentrates, and shows which of the tools you probably run have had the most trouble.
Where to go next
Check my AI plan
Describe what you are building. Get a verdict with percentages and the fewest changes that stop real attacks.
Test my AI
See whether your own assistant obeys an instruction hidden in a web page. Nothing you paste is stored or sent.
What's happening this week
The week's major incidents, news and new advisories, on one page with every source.
Best practices
Security practices keep attackers out of your AI systems. The agent security patterns are thirty controls for agents that can act, each linked to the practice behind it. Responsible AI controls keep the systems themselves fair, explainable and under human control. AI roles explains the new jobs and which part of the plan each owns. Education lists the training and certifications worth your time, independently chosen. Sources are linked on every card.
Grouped by where you apply them; the last line of each says how the curator helps.
Major incidents
Not every story, only the incidents that matter: real, publicly reported, and consequential, meaning a production breach, a mass exposure, or a flaw in something widely used. Each one has how it happened, the fix and the lesson, checked against the linked sources. Smaller and newer items live in until they earn a place here. Read them as a library, on a timeline, on the map, or by the industries they hit, race the attackers on the race clock, or follow them stage by stage along the kill chain.
Loading the map…
AI security in the news
Recent stories, newest first, summarized in our words with every outlet we read linked, so you can go to the original reporting. The tab lists where they come from.
Daily Today's breach puzzle: can you name the real incident in five clues?AI risk watch
Concerns about AI from different kinds of sources, kept apart on purpose: things AI systems have actually done, documented by a lab, a government or a named research team, and projected risks from authoritative reports. Every entry says who is speaking and links to them.
External sources
The outlets, research teams, official databases and lab pages this center is built from, plus independent voices (newsletters, podcasts, communities and bug bounty programs) that often carry a story first, then the frameworks and guidance it maps to. Sources with a feed can be added to any RSS reader in one go.
Removed from this directory in October 2026: The Hacker News. Earlier entries keep the links they were checked against.
AI models
Technology watch list
The AI software teams use most, and what has gone wrong with it. Usage figures are the vendors' and registries' own published numbers, each linked to where it can be checked. CVE counts come from this center's advisory feed and incidents from the incident library, so they move as entries are added; each card links to the full advisory record. A high count can mean heavy scrutiny as much as weak engineering, so read the incidents and advisories, not only the number.
Infrastructure
The physical side of AI: whether the services you depend on are up, where the major AI data centers are, the undersea cables that connect them, and the satellites that back them up, with a short video tour of the maps. What has gone wrong lately is in .
Service status
The official status pages of the AI services and developer platforms teams depend on, each opened and checked when it was added. When Claude, ChatGPT, GitHub or npm has an outage, or one model is degraded while the rest work, these pages are where the vendor confirms it, and most have a feed you can send to a team channel. The morning refresh records each vendor's own summary once a day and the cards show it; Check live status on a card opens that vendor's page, and nothing here reads a vendor page on your behalf.
Loading the map…
Video tour: AI data centers and undersea cables
A minute and a half, no sound: how to read the two infrastructure maps in this section. Where the major AI data centers are and how much power they draw, then the undersea cables that connect them, who owns each one, and where cables were cut. Pick a chapter to jump to it.
What it shows
The AI data center map sizes each site by its IT power and marks what is operating, being built or planned; filters narrow it by status, company and region. The cable map draws each cable between the places it lands, colored by owner (Google, Meta, Microsoft, Amazon or a carrier consortium), and marks every cut since 2024 with the report behind it.
Why it matters for AI security
Prompts, training data and model weights travel between a few very large buildings over a few cables. A cut, a regional outage or a tapped link reaches every AI service routed through it, so knowing where your vendors run, and over which routes, is part of planning for failure.
Loading the map…
Events
Conferences, summits, vendor keynotes, webinars and online demos that matter to AI security, from Google, OpenAI, Anthropic, Microsoft, NVIDIA, AWS, Meta and the security community. Each one says why it is worth your time and how to follow it without travelling. Dates were checked when added; confirm on the organizer's page.
Glossary
The words this center uses, in plain language. Each entry points to the tab that goes deeper.
About this center, and how the numbers are made
The AI Security Learning Center is free, public and independent: not funded by or affiliated with any AI lab, cloud provider or security vendor, with no sponsored content, no paid placements and no affiliate links. It is built and maintained by Eduardo P. Molina, with the rules below applied to every entry. Findings and ratings are for review, not proof of compromise.
What goes in
- Incidents: real, publicly reported incidents and disclosures, checked against the linked sources (researcher write-ups, vendor advisories, NVD, reputable news). Every claim links to where it came from. An entry says how it happened, the impact, the fix and the lessons, and names the industries of the victims only as the sources state them.
- Attribution: only what a named source publicly stated, never our own inference, with the source's own confidence level. When no source names a country, the entry says so.
- In the news: stories summarized in our words, with every outlet we read linked. Today's headlines come straight from the outlets' feeds in Sources, unedited.
- Known CVEs: published, fixed vulnerabilities with their advisory, affected and fixed versions, and a mark when CISA lists them as known exploited.
- Everything else (practices, weaknesses, frameworks, education, data centers, cables, satellites) cites its sources on the page; where a figure is ours, the page says "our rating" or "our estimate".
Severity ratings
An incident's severity is the published CVSS rating when one exists, and the entry names the source. When there is none, it is our own editorial rating, marked "our rating", on the same scale: critical (code execution, data theft or money lost at scale, or a flaw exploited in the wild), high (exploitable with real impact, limited by access or scope), medium (needs unusual conditions, or impact was contained), low (disclosure with no known victim and little practical impact). Advisories keep GitHub's own severity.
The threat index
The gauge at the top of the sections menu runs from 0 to 100 and is the average of four indicators, each from 0 to 100, read every morning off the incident library and the GitHub advisory snapshot. It is not a prediction: it says how much has gone wrong lately and how bad it was.
- More than usual: this week's incidents, weighted by severity (critical 3, high 2, medium 1, low 0.5), against a usual week of the previous twelve. A usual week scores 50, double the usual 75, half 25.
- How bad: the average severity of the last 14 days' incidents: all medium scores 20, all high 60, all critical 100.
- Real attacks: how many of the last 14 days' incidents hit someone (an attack carried out, victims in named countries, or a flaw that exposed people everywhere), rather than a disclosure with no known victim: one scores 25, three 58, six 82.
- New holes in the tools you run: critical and high GitHub advisories of the last 14 days for the tools on the Technology watch list; a critical counts as three highs, and a known exploited one three times more (one high scores 8, one critical 22).
Zones: Low below 25, Moderate 25 to 49, Elevated 50 to 74, High 75 and above. The gauge keeps 30 days of history; open it to see the four numbers behind today's value.
Industries hit
Each incident and story names the sectors of its victims or targets as the sources state them; one entry can name several, and a flaw in a tool that any company runs counts under every industry. Incidents with no named victim are listed apart and never guessed.
What is left out on purpose
- Government, military, intelligence and surveillance satellites, and attacks attributed to governments or countries, on the Satellites and space tab: only publicly known systems from technology companies are listed.
- Speculation: no entry is based on a rumor, a single unverified post, or our guess at who did it.
- Vendor marketing: a vendor's own write-up of its product is cited as a source only for facts about that product.
Updates and translations
The center is updated every morning: the headlines and the GitHub advisories are refreshed automatically, and the day's incidents, stories, events and tab changes are reviewed and added by hand, each marked New or Updated in What's new that day. Spanish is kept complete every day; French, Portuguese, German and Japanese get the headings and navigation daily and the full text weekly. Translations are made with AI from the English, which stays the reference; every translated page says so and links to the English.
Corrections
If an entry is wrong or a source is missing, every card has a link to copy: send it with what should change to the maintainer, and the fix is marked Updated in What's new on the day it lands.
Checklists and templates
Am I exposed?
Ten yes-or-no questions about how your team runs AI tools, agents and dependencies. Each one is a pattern behind real incidents in the library. It takes two minutes, and your answers stay on this page: nothing is sent or stored.
Want the fuller version? The security plan questionnaire asks eleven more detailed questions about how your team uses AI and turns the answers into a plan with roles, controls and a roadmap: see the sample plan, or answer it in a workspace.
AI security lab trial
All labs
Every hands-on tool in the center on one page. Each runs in your browser, and nothing you type or paste is sent anywhere.
In the Learning Center
In the other areas
Try it yourself
Ten hands-on tools. Name the day's real breach in five clues or fewer, spot a deepfake before it costs you, take a phishing test with an AI assistant reading along, see the text on a page that only the AI can see, find which real incidents could reach an agent set up like yours, see whether your own AI assistant obeys a hidden instruction, check the config files your agents obey, draw your AI risk portrait, play Spot the injection against the clock, and make the calls as CISO in three real attacks. Nothing you do or paste here is sent anywhere; the puzzle keeps your streak in this browser only.
What to look out for
What to act on now, what to watch for and what is coming, each item drawn from sourced entries in this center.
This week in AI security
Every alert, in detail
The digest above is the short read. Below it, every incident, story and advisory for the period you pick, updated every morning: incidents, major and medium only, from this center's library, already checked and rated; the news from In the news; and new GitHub-reviewed CVEs for the tools on the , with GitHub's own severity.
Preview AI Security Curator
Five things the curator does, shown on static examples. Nothing is uploaded or scanned on this page and nothing is stored; every result below is the curator's own check run on a fixed sample. A workspace does the same on your dependency files, repositories and agent configuration, adds OSV and Trivy, and re-checks every night.
The packages, images and AI components teams install most, each at a real version, with the verdict the built-in AI CVE feed and the default policy give it. The same version can be blocked at one release and approved at the next.
| Component | Verdict | Findings | Fix | What the curator does |
|---|
Vulnerabilities
Two kinds, side by side. Weaknesses are the patterns behind the incidents (how to detect each, whether it can be resolved and how to investigate), mapped to the OWASP Agentic and LLM Top 10, MITRE ATLAS, NIST AI RMF and ISO/IEC 42001. Known CVEs are the published, fixed flaws in the AI software and the everyday packages teams install; each CVE links to the weakness it is an instance of.
AI vulnerability database
Every weakness with how to detect it, whether it can be resolved, what to do when it cannot, and how to investigate. Mapped to the OWASP Agentic and LLM Top 10, MITRE ATLAS, NIST AI RMF and ISO/IEC 42001. Major weaknesses are the ones behind most real incidents; the others are narrower or lower impact, and still worth checking.
Every weakness has its sources. Open one to see them: the research, advisories and documentation it is based on, the incidents in this library where it happened, and the OWASP and MITRE ATLAS entries it maps to, linked under their official names. Nothing here is added without a source.
Known CVEs: AI ecosystem and everyday packages
Published, fixed vulnerabilities in model loaders, inference servers, agent frameworks, MCP tooling and coding agents, plus the packages every team installs: Django, Flask, PyYAML, Pillow, lodash, Express, Next.js, axios and more. Entries in CISA's Known Exploited Vulnerabilities catalog are marked.
| CVE | Package | Affected | Severity | Summary |
|---|