Common Types of Leaked Credentials (and the Fastest-Growing One)

The credential you’re picturing is probably wrong

Ask most engineers what a “leaked credential” looks like and you’ll get the same answer: an AWS access key, maybe a GitHub token. Both are real, common, and worth catching. But they’re also the easiest case — well-known formats, easy to pattern-match, easy to validate by just asking the provider’s API “is this key still live?”

The actual data tells a messier story, and it’s worth knowing about before you pick — or trust — a scanning tool.

The fastest-growing leak category isn’t cloud keys — it’s AI

According to GitGuardian’s State of Secrets Sprawl 2026 report, AI-service credential leaks — OpenAI, Anthropic, Hugging Face, and similar — grew 81% year over year, reaching over 1.27 million leaked secrets in 2025. Eight of the ten fastest-growing leak categories that year were tied to AI services, and LLM infrastructure tokens specifically leaked five times faster than core model-provider keys.

This tracks with how fast teams have adopted AI tooling: a new API key gets generated for every project, every prototype, every internal script calling a model — and, like any other credential, it gets pasted somewhere convenient. A Jira ticket documenting “here’s the key I used to test this,” a Confluence page with a working code sample, a comment on a bug report. None of that is unusual. It’s exactly how the previous generation of cloud keys leaked too — the pattern just moved to a newer category of secret.

Why “we validate our keys” isn’t the whole answer

A lot of scanning tools lean on validation: find something that looks like a key, then call the relevant provider’s API to confirm it’s actually live before flagging it. That’s a genuinely useful signal — it cuts down on false positives for well-known formats. But the same GitGuardian report found that 46% of critical secrets are missed by validation-only prioritization.

The reason is straightforward: validation only works for credential types with a provider API to check against. A private key, an internal service’s custom token format, a plain password= assignment in a config snippet — none of these can be “pinged” to see if they still work. A tool that only trusts what it can validate quietly ignores almost half of what actually matters. Severity needs to come from context — is this a paired, usable secret or a harmless identifier — not just from whether an API call succeeded.

What’s actually out there

Real leaks aren’t limited to one or two categories. Here’s the breadth of credential types that show up in practice across cloud infrastructure, developer tools, and everyday SaaS:

That last row — generic and cryptographic secrets — is exactly the category validation-only tools tend to miss, and it’s a big one: private keys and bare tokens don’t announce which service they belong to the way an sk_live_... Stripe key does.

The part nobody fixes on their own

Even when a leak is found, it often isn’t rotated. The same report found that nearly 70% of credentials confirmed valid in 2022 were still valid in January 2025 — and when retested a year later, the rate was still above 64%. A credential doesn’t get safer with age; it just sits there, waiting.

That’s an argument for detection that escalates, not just detection that logs a finding somewhere. A dashboard nobody checks doesn’t rotate a key. A Jira incident assigned to someone, with a due date, has a much better chance of actually getting fixed.

Where this leaves you

If you’re evaluating how well-covered your team actually is, it’s worth checking two things: whether your scanning covers credential types beyond the easily-validated ones (private keys, internal tokens, AI service keys), and whether a real finding turns into something a person is actually responsible for closing out.

Secret Sentinel scans Confluence and Jira for 50+ credential types — including OpenAI, Anthropic, Groq, and Hugging Face keys alongside the more familiar cloud, database, and payment provider formats — classifies findings by real exploitability rather than validation alone, and can automatically open a tracked Jira incident for anything high-risk.

Install Secret Sentinel for Confluence · Install Secret Sentinel for Jira