For years, AI-powered attacks were a warning about the future. In October 2026 they became news. AI agents found a flaw that human reviewers had missed for over a decade, and autonomous agents broke into real companies and banks.
This post covers what happened, what the attacks have in common, and how Guardian's Samurai Security lets you find the same weaknesses in your own MCP servers before someone else does.
What happened in October
PHP2Shell: an AI agent finds a ten-year-old flaw
According to TIME, a Wiz researcher, Sagi Tzadik, asked an AI agent whether a flaw in WordPress could work against Drupal. The agent reported that it did not, but along the way it had found a flaw in PHP that "had gone undetected for more than a decade."
Calif, working with OpenAI, had also used AI to find a related flaw. Then the race began. Word of the vulnerabilities began circulating online two days before the patch was due, and Calif mistakenly posted a video showing its exploit at work. TIME reports that Cloudflare counted thousands of attempts before the public announcement, and the number later reached millions. While teams installed the update, many webpages, including government agencies and parts of Wikipedia, were taken down as a precaution. TIME says as much as 70% of the web depends on PHP.
The Hacker News names the flaws as CVE-2026-109618 and CVE-2026-109619, and reports that threat actors were seen using agentic AI platforms for activity after they got in. It also notes a correction: Unit 42 withdrew its earlier assessment that an attack on a European government agency involved the PHP2Shell flaws.
One attacker, about 100 companies
According to Reuters, as reported by Fortune, a hacker set autonomous AI agents loose on websites starting September 25. The agents, using DeepSeek and Claude Code models, tried more than 1,200 websites and breached about 100 companies, including Deere, SharkNinja and Cummins. The tokens cost about $5,000, roughly $50 per breach.
South Korean banks
According to Reuters, also reported by Fortune, attackers likely based in China used Artex, a Chinese state-developed AI tool, together with Claude Code to attack South Korean banks. The Wall Street Journal reported that they stole data from an estimated 68,000 people.
OpenCode
The same Fortune article describes an attack on OpenCode that went on nonstop for about 48 hours. The company said attackers broke into accounts affecting 1.4% of its users, and Dealroom reported that some users lost as much as $1,700.
The pattern
Different targets, same story. Agents find weaknesses and try them faster than most teams can patch.
The weaknesses agents go after are rarely exotic. For an MCP server, which is the door between AI agents and your tools and data, they come down to a short list:
- missing or weak authentication
- forged or damaged tokens
- unsafe OAuth settings
- missing Origin checks
- secrets leaking out of tool responses
- prompt injection hidden in tool responses
- path traversal, where a request reaches files it shouldn't
If an attacker's agent is going to try these, the safest move is to try them first, on your own server, with permission.
How Samurai Security checks for the same gaps
Samurai Security runs an authorized live security test against your own MCP server. Every run checks the server and each of its tools against seven attack classes:
- Authentication bypass: calls each read-only tool with no credentials and reports any call the server should have refused.
- Token forgery: calls each read-only tool with the saved credential damaged into an inert value and checks that the server rejects it.
- OAuth misconfiguration: checks for a proper authentication challenge and that the URLs in your OAuth metadata are safe and match where they are published.
- DNS rebinding: sends a request with a foreign Origin header and checks that the server rejects it, as the MCP transport spec requires.
- Secret exposure: scans real tool responses for API keys, tokens and private keys. Only the secret type and a short prefix are kept.
- Prompt injection: scans real tool responses for text that tells an AI agent to ignore its instructions or leak credentials.
- Path traversal: sends a ../ traversal path to tools that take a file or path and checks that no system file comes back.
You stay in control. Before anything runs, you get a security plan you can review: every test grouped by attack class, what it applies to, exactly what is sent and whether it sends an attack payload. When the run ends, a results matrix shows your server and its tools down the side and the attack classes across the top, so you can see where you are exposed at a glance. Each failure comes with a finding and a suggested fix.
Test every environment
- Production servers, from the web app. Save the server, open Samurai Security and run the test. The web app tests saved servers marked as production.
- Staging and preview deployments of development branches, from a terminal or a CI pipeline. Run
guardian security-test SERVER --fail-on high, where SERVER is the name of the saved server. The command fails the build if a high-severity problem turns up, so a risky change is caught before it merges. - A server on your own machine, before you push. Use
guardian test, against a loopback address or with--stdio, to check how it behaves while you build.
One thing to be clear about: Guardian's cloud runs the security battery against a saved http server that it can reach. It does not reach into your laptop or your local network. That is why a deployed staging or preview address is the right place for it, and why guardian test is the tool for a server that only exists locally.
A few requirements apply. Only an Admin or a Security member can run the security test, using a server whose credentials they saved themselves. Your plan needs to be Standard or above, and an Admin must have accepted the terms in the web app. In a pipeline, use a run token from an Admin or Security member, give it a short expiry, and scope it to the jobs that need it.
Close the gaps: a short checklist
- Require sign-in on every MCP endpoint, and reject missing or damaged tokens.
- Check that your OAuth metadata URLs are safe, and reject requests with a foreign Origin header.
- Keep secrets out of tool responses, and treat text coming back from a tool as untrusted.
- Make sure tools that take file paths can't reach outside their folder.
- Run the security test on production and on every staging or preview deployment, and re-run it after each change.
Start before the next headline
Learn more on the Samurai Security page, or create a free account to get started.
One reminder: only test servers you own or are authorized to test.
Photo by Tasha Kostyuk on Unsplash.
