AI governance, risk and continuity6 min read

AI Agent Credential Leaks: A Business Continuity Checklist

What the Grok Build and Claude Code exposures mean for Singapore companies: why a leaked key stops the business, what buyers will ask, and a checklist.

Haojun See
Haojun See

Founder & CEO, On The Ground

Updated 15 November 2026

What happened with Grok Build and Claude Code in November 2026

Two reports in one week showed that AI agents now hold company secrets, and that default settings decide whether those secrets stay private. Neither was a hack. Both came from settings people accepted without a second look. Grok Build. Security firm Claroc found that 35,578 of 37,571 live apps built with SpaceXSI's Grok Build (about 95%) opened for anyone, without signing in. The access setting was preselected as "Anyone with the link", and app addresses leaked into public DNS records when the apps were visited (The Next Web). Claroc estimated 650 to 800 apps held private data, including medical records, employee details and plaintext passwords. Twenty-four exposed apps belonged to SpaceXSI and X staff, including SpaceXSI's own sales pipeline (The Decoder). SpaceXSI's response, as reported, is that apps are private until their creator shares them. Claude Code. Security firm Apiiro reported on 5 November 2026 that choosing "Yes, and don't ask again" writes the approved command into a project file, .claude/settings.local.json. If that command contained a password or API key, the secret sits there in plaintext. Apiiro examined 104,185 public copies of the file on GitHub and found 1,122 holding 1,379 distinct secrets (iTWire). It also reported that popular open-source secret scanners missed a large share of them. Be precise about the second case: Claude Code did not publish the files. The file is meant to stay on the developer's machine, and a Claude Code team member said Claude Code adds it to .gitignore when it creates the file. Developers committed it to public repositories anyway. To a client whose data leaked, that distinction matters little.

Why a leaked AI agent credential is a business continuity problem

A leaked credential stops the business, not just the IT team. Once a key is exposed you must rotate it, and everything that depends on it breaks until it is replaced. • Forced downtime. Rotating a payment, database or cloud key takes systems offline. If nobody knows which apps use which key, the outage runs for days. • Hidden dependencies. Apps built in an afternoon with an AI builder often become part of daily operations without ever appearing in an IT inventory. If a vendor pulls public apps, those workflows vanish without notice. • Regulatory clock. Under the PDPA, a notifiable data breach must be reported to the PDPC within three calendar days of assessing it as notifiable. That clock runs while you are still restoring service. • Client trust. Clients judge you on whether their data was safe, not on whose default setting caused the leak.

What procurement teams will ask about AI agents

Expect vendor questionnaires to add a section on AI agents. If you sell to large companies or government agencies, prepare answers to these questions now: • Which AI agents and app builders can touch our data? Show a named list with an owner for each tool. • Where do agents store credentials and approvals? Secrets belong in a vault or environment variables, never in config files or chat logs. • What is the default sharing setting for anything an agent publishes? It should be private, with sign-off needed to share. • Are agent config files kept out of version control? Show .gitignore rules plus a secret scan on every commit. • How fast can you rotate a key, and what breaks when you do? Show a tested rotation runbook with a stated time. • How would you know if an agent-built app went public? Show a regular audit of published links. The scanner results are a warning for anyone who answers "we run a scanner": by Apiiro's account, standard tools missed many of the leaked files. Buyers will start asking how you know your controls work, not just whether you have them.

Checklist: what to do this week

Most of this takes an afternoon and needs no new software. • List every AI agent, coding assistant and app builder in use, including ones staff signed up for themselves. • Open each app builder's sharing settings. Set the default to private and check every published link. • Search your repositories for .claude/settings.local.json and similar agent config files, and add them to .gitignore. • Rotate any key that has ever appeared in a config file, a chat with an agent, or a pasted command.

Checklist: what to fix this quarter

Make the fixes stick by changing where secrets live and testing your recovery. • Move secrets into a password manager or secret store. Agents read them from environment variables, so the value never appears in a command. • Write a one-page key rotation runbook: which key, which systems use it, who rotates it, how long it takes. Test it once. • Run two secret scanners in your build pipeline, not one. • Map which daily workflows depend on an agent-built app, and name a fallback for each if the vendor removes it. • Treat "don't ask again" and "share with anyone" as decisions that need a second look, and add AI agents to your business continuity plan.

Frequently asked questions

Did Claude Code leak passwords to GitHub?

Not directly. Choosing "Yes, and don't ask again" saves the approved command in .claude/settings.local.json, which can include a password or key. The file is meant to stay local, but developers committed copies to public GitHub repositories. Apiiro found 1,122 such files holding 1,379 secrets. Keep the file out of version control and rotate any exposed key.

How many Grok Build apps were publicly accessible?

Security firm Claroc found that 35,578 of 37,571 live Grok Build apps, about 95%, opened for anyone without signing in. The default access setting was "Anyone with the link", and app addresses leaked into DNS records. Claroc estimated 650 to 800 held private data. Twenty-four belonged to SpaceXSI and X staff.

Why is a leaked API key a business continuity risk?

Because the fix causes downtime. Once a key leaks you must rotate it, and every system using that key stops until it is updated. Without a list of which apps use which key, recovery can take days. In Singapore, a notifiable breach involving personal data must also be reported to the PDPC within three calendar days.

How should companies store credentials used by AI agents?

Keep credentials in a password manager or secret store and let agents read them from environment variables, so the value never appears in a command, config file or chat log. Keep agent config files out of version control, run more than one secret scanner, and keep a tested runbook for rotating each key.

Train your team on this

We teach teams to use Claude on their own work, and set up a company AI for them. Start with a free 1-hour preview, or see the full package.

Want to Apply This to Your Business?

We're a Singapore AI development and automation agency. Let's discuss how we can help solve your specific challenges.