Learn

Credential exposure

Are Leaked Passwords Too Old to Matter?

A credential exposure report published in early November 2026 found that 97 per cent of the credentials it examined came from old breach dumps and credential lists, and under 3 per cent came straight from infostealer malware. Press coverage reduced that to a headline about leaked logins being old. It is a fair reading of the data and it raises the obvious question for anyone paying for monitoring. If the material is years old, has it not all been changed by now? The honest answer comes from a study that checked roughly 225,000 university usernames against breach data without touching a live account. The median guessed password had still been valid for 6.2 years.

What the November 2026 report found about leaked passwords and their age

Outpost24 published its 2026 Credential Exposure Benchmark Report on 3 November 2026. The numbers are specific, which is to its credit. Its researchers analysed 31.5 million unique exposed credentials linked to 4,824 organisations across five industries in the United States and Europe, first detected between 1 September 2025 and 1 September 2026. The headline, in the company's own words, is that "72% of those organizations had staff or customer credentials identified in infostealer malware logs during that period."

Read that sentence carefully, because it is more careful than most of its coverage. "Those organisations" means the 4,824 the dataset was built around, every one of which had exposed credentials by construction. So the figure is 72 per cent of organisations already known to have an exposure, not 72 per cent of businesses. Outpost24 included the qualifier. Two of the five outlets that covered the report dropped it.

The split underneath the headline is the part worth your time. Around 30.7 million of the credentials, roughly 97 per cent, came from breach dumps and credential lists, the large collections described in what is a combolist. The remaining 844,000, under 3 per cent, came directly from infostealer logs, which the report says are far more likely to reflect recent theft. One outlet ran that as "most leaked logins are old", which is a fair reading.

It is worth noting that publishing the under 3 per cent figure cuts against Outpost24's own commercial interest, since the newer and more alarming category turns out to be the small one. Vendors do not usually volunteer that.

What the report does not measure, including age

The report publishes no age distribution for the credentials. What it publishes is a classification by source, breach dumps treated as older material and infostealer logs as newer. That is a reasonable proxy, but a proxy for freshness is not a measurement of age, and three of the outlets covering it said plainly that no age or staleness figures are given.

There is a subtler problem in the window itself. The credentials were "first detected" between those two dates. That is a window of discovery, not of theft. A password stolen in 2018 and reposted in a new list in March 2026 is first detected inside the window and counts as such. Nothing in the method can separate the two cases, and the report does not claim to.

This is not a flaw unique to Outpost24. It is a property of the material. A September 2026 academic study of infostealer logs states it directly: "A log records when we collected it, not when the credentials were stolen or when they began circulating," and elsewhere that "the logs do not record when a credential was saved." The timestamps everyone wants simply are not in the data. Which means neither a vendor nor a critic can prove the feed is fresh or stale. Anyone who tells you otherwise is inferring.

What the report does state is its sample size, its sector list, its geography and its exact dates. What it does not state anywhere we could read ungated is how the 4,824 organisations were selected, whether they are customers or a wider scan, how a credential was counted, what the word "unique" was deduplicated on, or whether any credential was checked to see if it still worked. Four of the five outlets noted the absence of a stated method independently.

One point in the report's favour, and it is a real one. Its headline is expressed at organisation level, and organisation-level percentages are far more robust to deduplication choices than credential counts, because they need only one hit per organisation. The 72 per cent survives most reasonable arguments about counting. The 31.5 million does not.

The research answer on how long a leaked password stays valid

If 97 per cent of this material is old, the fair question is whether it still opens anything. The problem is that testing a leaked password against a live account is unauthorised access, so the honest research is scarce.

One study solved it properly. Researchers at Carnegie Mellon, the Max Planck Institute for Security and Privacy and the University of Chicago took roughly 225,000 university usernames, checked them against more than 450 individual breaches and 12 breach compilations, and had the institution's own IT staff verify guesses against its internal password history database. No live account was touched. The results, published in July 2023: the median guessed password had remained valid for 6.2 years, the longest for 19.8 years, and for 35.5 per cent of the accounts where any password was guessed correctly, the guess was the person's current password. Guesses succeeded for 32.0 per cent of accounts whose university email appeared in a breach, against 6.5 per cent of accounts with any potential match at all.

So "old" and "dead" are different claims, and the gap between them is years. That is the answer to the staleness objection, and it is the strongest single finding in this article.

Now the part that cuts the other way, because it should be said. A study published in June 2025 by researchers in Ljubljana ran 39 honeypots across 14 services from March 2023 to March 2024 and compared what attackers actually tried with a dataset reduced from over 27 billion leaked records to more than 4 billion unique ones. Unique attacking credential pairs came to less than 0.01 per cent of the unique accounts in that leaked data. The paper's own conclusion is that "this overlap was relatively limited, meaning that attackers are not simply reusing leaked credentials en masse." What the honeypots mostly saw was common word lists and default credential lists.

Hold both findings at once, because they are not contradictory. Leaked passwords stay valid for years. Attackers mostly do not bother spraying them indiscriminately. Which means the risk is not volume, it is targeting: the leaked credential matters when someone chooses to use it against you specifically.

Two provider figures sit alongside this. Microsoft documents that its leaked credentials detection validates breach material against a tenant's current valid password hashes and emits a detection only on a confirmed match, calling it verified credential exposure rather than a heuristic signal. It publishes no match rate. Cloudflare, across about 30 million properties from September to November 2024, reported that 41 per cent of successful logins involved compromised passwords. Google's own older telemetry, from 2017, put the validity of breach-exposed credential pairs at 7 per cent due to reuse. Reuse itself is the mechanism, and the best-sized study on it, from Virginia Tech, found 38 per cent of users had at least one identical password pair across 107 leaks, with a further fifth modifying an existing password, and 30 per cent of those modifications guessable within ten attempts. That last figure is the reason an old password still has value to an attacker: it reveals a pattern, and patterns get modified rather than abandoned. The mechanics of what is actually in these files are covered in plaintext, hashed and salted passwords, and how the guessing works in what is credential stuffing.

Where this criticism lands on us too

The strongest argument against credential exposure monitoring is not that the data is stale. It is that monitoring is structurally late. You only see what an attacker chose to publish, which means the service cannot honestly promise early warning, and the same budget spent on single sign on with phishing-resistant authentication removes the risk rather than reporting it.

Four parts of that apply to SphereTI exactly as they apply to Outpost24, and we are not going to pretend otherwise.

We are downstream of whatever gets released. Our visibility starts when material reaches a channel we collect from. If it never circulates, we never see it.

Our dataset mixes new, old, duplicated and unverifiable records, for the same reason everyone's does. The timestamps that would let us sort them do not exist in the source material.

An alert does not know the context. It cannot tell you whether that person still works for you, whether the password was rotated last year, or whether the account has multi-factor authentication on. You know those things. We do not.

Our own denominators are selected on exposure too. Any figure we publish about "organisations with exposure" is drawn from organisations already present in breach data, which is the same structural property as the 72 per cent discussed above.

What we would say in reply is narrower than a sales pitch. An exposure record is a targeting signal. It tells you which of your people appear in material an attacker can buy, and when it came from an infostealer log, it tells you a device was compromised and that the password was the least interesting thing taken from it. That is why the right response to a detection is to revoke sessions and treat the device as compromised, not to reset a password and stop. Whether that signal is worth paying for depends on something the data cannot tell you: whether anyone at your company is in a position to act on it within a day. If nobody is, buy the capacity to act first. The guidance in evaluating dark web monitoring providers is a reasonable place to start, including on us.

Common questions

How long does a leaked password stay valid?

The best available answer comes from researchers at Carnegie Mellon, the Max Planck Institute for Security and Privacy and the University of Chicago, who checked about 225,000 university usernames against more than 450 individual breaches and 12 breach compilations. Guesses were verified by the institution's own IT staff against its password history database, so no live account was ever tested. The median time a guessed password had remained valid was 6.2 years, and the longest was 19.8 years. For 35.5 per cent of the accounts where they guessed any password correctly, the guess was the person's current password.

If the data is mostly old, why does monitoring cost money?

Because the age of the record is not the same as the state of the account, and because an exposure record tells you something the password alone does not. A credential from an infostealer log means a specific device was infected, which means everything else saved on that device left at the same time. The defensible claim for monitoring is narrow: it is a targeting signal that tells you which of your people and devices an attacker already has material on. It is not a defence against generic password guessing, and anyone selling it as one is overselling it.

What does 72 per cent of organisations actually mean in that report?

It means 72 per cent of the 4,824 organisations the report's dataset was built around, all of which already had exposed credentials linked to them. Outpost24's own wording is careful about this and says "72% of those organizations". Two of the five outlets that covered it dropped the qualifier and wrote "72% of organizations", which a reader could easily take as 72 per cent of businesses generally. That is not what the report says, and the difference matters when you are deciding whether a number applies to you.

Do attackers really try leaked passwords against accounts?

Less than the marketing suggests, according to the one study that measured real attack traffic. Researchers in Ljubljana ran 39 honeypots across 14 services for a year and compared the logins attempted against them with a dataset of over 4 billion unique leaked records. Unique attacking credential pairs amounted to less than 0.01 per cent of the unique accounts in the leaked data, and the paper concludes that attackers are not simply reusing leaked credentials at scale. What attackers mostly sprayed were common word lists and default credential lists.

Then which evidence says leaked credentials are still live?

Two sources, both from providers rather than academics. Microsoft documents that its leaked credentials detection validates breach material against the tenant's current valid password hashes and emits a detection only when a confirmed match is found, which it describes as verified credential exposure rather than a heuristic signal. Microsoft publishes the mechanism but no match rate. Cloudflare, looking at about 30 million properties between September and November 2024, reported that 41 per cent of successful logins involved compromised passwords. Neither figure tells you what share of any given leaked dataset is still live.

What should we actually do with an old exposure record?

Treat it as information about a device and a person, not only about a password. If the record came from an infostealer log, assume the device was compromised and that session cookies, saved logins and any keys on it left too, then revoke sessions rather than only resetting the password. If it came from a breach dump, check whether that person reuses patterns, make sure the account has phishing-resistant sign-in, and move on. The one thing not to do is reset the password, feel finished, and leave the live session untouched.

Check your own exposure

Enter a work email and SphereTI checks it against known breaches and stealer logs, then shows you which credentials have been exposed. Free, and it never asks for your password.

Get your free report

Keep reading