AI and Cybersecurity Learning and Practice Center Mission control AI ExplainedPreview Test automationPreview AI DeliveryPreview Open-source reposPreview

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.