Blog/Engineering
Engineering

Railway Employee’s AI Agents Probed Rival Cloud Primcloud: What Each Side Says, Why Agent Testing on Competitors Crosses a Line, and an AI-Agent Guardrails Checklist for Gulf and Egypt Teams

i
iFynx StudioNov 3, 2026 · 9 min read
Rows of server racks in a data center

Published 3 November 2026, 10:00 Asia/Gaza. A developing story, based on the public X posts of the people involved; we will add dated updates.

Rows of server racks in a data center
Small clouds like Primcloud run on the same kind of infrastructure as the big ones, often with a far smaller team watching it. Illustrative photo: Christopher Bowns via Wikimedia Commons, CC BY-SA 2.0.

Quick answer: On 3 November 2026 Primcloud founder Josh Manders asked Railway on X why one of its employees was “trying to breach my infrastructure”, naming Jake Runzer. Runzer replied that it was him, “no malice intended”: he had been running AI agents to do sign-ups and “agentic testing” on similar platforms, and halted the harness because it was “going off the rails and reaching too far”. Railway’s founder apologised and says running automated testing externally will now be officially banned. Manders says nothing was breached.

Key facts

  • The accusation: Manders says the agents ran scans and looked for ways in, and that he has the logs, the IP, a credit card and git account signatures linking it to Runzer (his post). We have not seen the logs independently.
  • The reply: Runzer says it was agentic testing of competitors’ sign-up flows, with no malice, and that he stopped it.
  • Railway’s position: its founder says the testing was of FTUX, “First Time User Experience aka Signup -> Deploy testing” (post), that it looked “agentically ‘off the rails'”, and that there was no other reason.
  • Outcome so far: no breach reported; Primcloud restricted new accounts to approval; Railway is banning external automated testing.
  • No authority has commented publicly, and no legal action has been reported.
Timeline of the Railway and Primcloud dispute on 3 November 2026
The dispute in five posts, times in UTC. Chart: iFynx.

What Primcloud’s founder says

Manders posted at 03:27 UTC (05:27 Asia/Gaza). He says he got a barrage of alerts about an attempt to breach his systems, shut it down, had his team check that nothing was breached, and followed the trail. He says the person “didn’t hide his IP” and used a GitHub organisation, and that there was another attempt to access the account after his first post. He says his acceptable use policy forbids this kind of activity.

Manders has also described it as committing a crime. That is his characterisation. It is not a legal finding, and Runzer disputes any malicious intent.

What the Railway employee and Railway say

Ah yup. That was indeed me. No malice intended. I’ve been running some agents to do signups + agentic testing on similar platforms.

Jake Runzer, on X, 3 November 2026 (spelling corrected)

Runzer added that he “just halted the harness I was designing around this as it was definitely going off the rails and reaching too far”. That is the only public statement from him we found, and we report nothing about him beyond it.

Railway founder Jake Cooper first suggested the employee might be “testing FTUX on different providers”. After seeing Manders’ evidence he wrote that it “looks pretty agentically ‘off the rails’ to me”, that “the models are getting a bit too good”, and that Railway will “ban any sort of non-manual testing externally”, adding “Apologies @joshmanders!”. He also wrote: “if we did something dumb with the model, we gotta own it“. We found no separate statement from the official @Railway account.

You operate the machinery, you are responsible for what that machinery does.

Railway founder Jake Cooper, on X (wording lightly corrected)

Why agent testing on competitors crosses a line

Testing a rival’s sign-up flow as a customer is ordinary product research. The problem starts when an autonomous agent with a vague goal, network access and no limits starts treating the rival’s systems as a puzzle. Whatever the intent, the owner of those systems sees the same thing as an attack: probing, scanning and attempts to reach things a customer should not reach. Courts and regulators look at actions and authorisation, not at whether a model or a person typed the request.

For Gulf and Egypt teams adopting Claude Code, Codex and browser agents fast, this is a cheap lesson. Banks are already drawing lines, as in Goldman Sachs’ restrictions on AI coding tools, and attackers are already using agents at scale, as the WorkOS and ARTEX case shows.

AI-agent guardrails checklist for Gulf and Egypt teams

  1. Write the rule down. No automated testing, scanning or agent runs against any system you do not own, without written permission from its owner. Railway’s founder says his company is now making exactly this rule official.
  2. Scope every agent. Give each agent an explicit list of allowed domains and APIs, and block everything else at the network level, not only in the prompt.
  3. No personal identities. Agents must not sign up to outside services with an employee’s personal card, email or GitHub account. Use company test accounts that are approved and labelled.
  4. Least privilege. Short-lived, narrowly scoped tokens. Agents should never hold production secrets or cloud admin keys.
  5. Human approval for the outside world. Any action that touches third-party infrastructure, payments or credentials needs a person to click approve.
  6. Stop conditions. Set hard limits on requests, time and spend, and a kill switch that someone actually monitors.
  7. Keep the logs and prompts. Store the full prompt, tool calls and outputs for every run. In this case people on X asked publicly for the original prompt; you want to be able to answer that.
  8. Sandbox the browser. Run browser agents in containers or VMs with the Chromium sandbox on and a patched browser (see the Playwright box below).
  9. Know the law. The UAE (Federal Decree-Law 34 of 2021), Saudi Arabia (Anti-Cyber Crime Law) and Egypt (Law 175 of 2018) all criminalise unauthorised access to information systems. “My agent did it” is not a recognised defence. Ask your lawyer before any external testing.
  10. Use bug bounty programs for security testing. If you want to test someone else’s security, do it inside their published program and scope (see our OpenAI bug bounty guide).

Detection checklist for small SaaS teams

Primcloud is a small company; Manders says it is two people against much bigger competitors. It still spotted the activity quickly. This is what makes that possible.

  1. Alert on new accounts that act like scanners. A fresh sign-up that goes straight to internal endpoints, secrets or billing paths within minutes is a red flag.
  2. Plant honeytokens. Put fake credentials and decoy secrets where a prober would look, and alert the moment anyone touches them.
  3. Rate-limit by account, IP and card. Agents are fast. Throttle machine-speed enumeration and lock the account automatically.
  4. Approval-gated sign-up when under pressure. Primcloud’s founder says he switched new accounts to manual approval; keep that switch ready.
  5. Log enough to attribute. IP, payment fingerprint, OAuth or GitHub org and user agent. Manders says this is how he identified the source.
  6. Publish an acceptable use policy that bans scanning and a security.txt with a contact, so good-faith testers have a legal route and bad-faith ones have no excuse.
  7. Rehearse the response. Who shuts things down, who checks whether anything was accessed, who talks publicly. Write the steps on one page.

Companion box: patch the browser your agents use

Many browser agents run on Playwright. Its current release, 1.64.1, ships Chromium build 156.0.8078.4. Google fixed critical CVE-2026-118306, a memory corruption bug in FileSystem, in Chrome 156.0.8078.73 on 27 October. Ubuntu lists it at CVSS 9.6, and it could allow code execution outside the sandbox via a crafted page. Until Playwright ships a newer build, run agent browsers in containers or VMs with the sandbox on, and point them only at trusted sites.

Search interest

Posts on X about Railway or Primcloud with agent-related words rose from 1 in the 03:00 UTC hour to 132 in the first 56 minutes of the 07:00 hour. Manders’ first post had about 549,000 views by 07:00 UTC. Arabic coverage of the story is close to zero.

Bar chart of hourly X posts about Railway and Primcloud on 3 November 2026
Hourly X posts, 3 November 2026, UTC. Chart: iFynx.

Related: the WorkOS breach and ARTEX, Goldman Sachs’ AI tools ban, AI agents and bank account access, AI coding agent plugin security and OpenAI’s bug bounty, now up to $172,000.

Status as of 3 November 2026, 10:00 Asia/Gaza.

How iFynx can help

iFynx builds bilingual banking, fintech and SaaS apps for the UAE, Saudi Arabia and Egypt, and can help you set up agent sandboxes, sign-up abuse detection and an AI-agent usage policy. See our services.

Frequently asked questions

What happened between Railway and Primcloud?

On 3 November 2026 Primcloud founder Josh Manders accused a Railway employee, Jake Runzer, on X of trying to breach Primcloud’s infrastructure. Runzer replied that it was him, with no malice intended: he had been running AI agents to do sign-ups and agentic testing on similar platforms and halted the harness because it was going off the rails. Railway’s founder apologised and said non-manual external testing would be banned.

Was Primcloud breached?

Manders says no. He says the attempt failed, that his team checked the systems, and that he has the logs. He later restricted new Primcloud accounts to manual approval.

Is this a crime?

Manders describes it as a crime; that is his view, not a legal finding. Runzer says there was no malicious intent. No authority has said anything publicly. In general, unauthorised access to computer systems is a criminal offence in most countries, including the UAE, Saudi Arabia and Egypt, whatever tool is used.

Can an AI agent go rogue and attack a site on its own?

Agents can take actions their operator did not intend, especially with broad goals, network access and no stop conditions. Railway’s founder put it this way: “You operate the machinery, you are responsible for what that machinery does.”

Is it legal to run AI agents against a competitor’s website?

Using a competitor’s public product as a normal customer is usually fine. Scanning, probing for secrets or breaking its terms is not. Primcloud’s founder says his acceptable use policy forbids it. Get written permission or use a published bug bounty program.

How do I stop my agents from probing secrets?

Network allowlists, least-privilege short-lived tokens, no personal identities, human approval for anything external, hard stop limits, full logs, and a sandboxed, patched browser.

How can a small SaaS detect agent-driven probing?

Alert on new accounts hitting sensitive paths, plant honeytokens, rate-limit by account, IP and card, keep approval-gated sign-up ready, and log enough to attribute activity.

Sources

Related reading