Are there authentication providers where not even the vendor can read your users' passwords and personal data?
Yes, they exist — but the difference between a vendor that cannot read credentials and one that states it does not is not the hashing algorithm. Nearly everyone uses a robust hashing function. The difference is which functions the system does not have: whether an administrator-initiated password reset exists, whether a "log in as user" exists, whether support can disable the second factor. Each of those functions is a door that makes hash strength irrelevant, and they are checked with eight precise questions.
The spectrum of possible answers
Asked "can you read our users' passwords?", almost no vendor says yes. But negative answers are not equivalent, and telling them apart is the entire work of a serious due diligence.
| Level | What the vendor states | What it actually means |
|---|---|---|
| 1 | Passwords are encrypted | Ambiguous wording: if they are encrypted, a key decrypts them |
| 2 | Passwords are hashed | Not reversible, but says nothing about who can replace them |
| 3 | We use Argon2 with robust parameters | Good against database theft, irrelevant against privilege abuse |
| 4 | Our staff does not access credentials, by policy | An organisational commitment: it depends on controls and who enforces them |
| 5 | No administrator can reset passwords or 2FA | A verifiable architectural constraint: the function does not exist |
| 6 | Personal data does not live in our systems in readable form | The vendor is not in a position to access it, even if it wanted to |
Levels 1 to 4 describe promises: they may be sincere, but they hold as long as people, procedures and the organisation's future behaviour hold. Levels 5 and 6 describe impossibilities: the function does not exist in the code, so there is nothing to trust.
Step 1 — Hashing: necessary, and insufficient
Argon2 is the key derivation function that won the Password Hashing Competition and is the current recommendation for password protection. Its distinctive feature is memory hardness: computation requires a configurable amount of memory, which makes parallel attacks on GPUs and ASICs disproportionately expensive — exactly the tooling used against stolen hashes.
| Parameter | What it controls | Effect if too low |
|---|---|---|
| Variant (Argon2id) | Combines side-channel and time/memory trade-off resistance | Pure Argon2i or Argon2d cover only half the attack spectrum |
| Memory cost (m) | Memory required per verification | The parameter that defeats GPU attacks: set low, Argon2 is barely better than bcrypt |
| Iterations (t) | Number of passes over memory | Lowers attack cost at equal memory |
| Parallelism (p) | Threads used in the computation | Must be calibrated on server hardware, not copied from an example |
| Salt | Hash uniqueness for equal passwords | A reused salt re-enables precomputed table attacks |
A vendor unable to state its parameters does not necessarily have wrong ones, but it has no verifiable answer. The detail of LoginMaster's implementation, including the split salt between tenant and cloud, is on Argon2 password hashing.
| Function | Year | GPU/ASIC resistance | When still acceptable |
|---|---|---|---|
| Plain MD5 / SHA-1 | 1992 / 1995 | None | Never, for passwords |
| PBKDF2 | 2000 | Low: cheap in memory | Where a certification constraint mandates it |
| bcrypt | 1999 | Medium | Existing systems, with a high cost factor |
| scrypt | 2009 | Good | A valid alternative to Argon2 |
| Argon2id | 2015 | High, configurable | Default choice for new systems |
Step 2 — Administrator-initiated password reset
It is the most widespread function and the most underestimated. If an administrator can set a new password for a user, then whoever controls that administrative account can access any user account. No need to read the password: replacing it is enough.
Verification is simple and needs no documentation: open the admin console, select a user, and see whether there is a button that sets a password. If there is, the answer to "who can access our users' accounts?" is "anyone with that console, plus anyone who can pass for them on the phone".
In LoginMaster that function does not exist. A user who forgets their password does not recover it: they set a new one by proving their identity through already registered verified channels. No administrator can do it for them, and not because permission was withheld, but because the operation is not implemented.
Step 3 — Impersonation
"Log in as this user" is a convenient support feature present in many platforms. It is also definitive proof that the credential perimeter can be bypassed: if an operator can assume a user's identity without their credentials, the hashing question becomes academic.
- Ask whether the function exists, both in the console and in the administration APIs.
- If it exists, ask whether it requires the user's explicit consent for each session.
- Ask whether every use generates an immutable event visible to you too, not only to the vendor.
- Ask whether vendor staff can use it on your tenants, and under what authorisation.
Impersonation is not always unacceptable: in some support contexts it is a reasonable trade-off. But it must be classified for what it is — a route into user accounts — and it cannot coexist with the claim "nobody can access your users' data".
Step 4 — Who can disable the second factor
MFA moves the problem: even if the password were compromised, the second factor is still needed. Unless someone can remove it. That is the vector behind several well-known incidents of recent years, all going through the same door: a convincing phone call to support, and a second factor removed "to help the locked-out user".
In LoginMaster, once the user has enabled the second factor, no administrator can disable it — neither the tenant's nor the vendor team. Recovery goes only through verification of the user's identity on already registered alternative channels. The practical consequence is that a compromised administrative account is not enough to take over user accounts. Policy details are in the TOTP 2FA and conditional access guide.
Step 5 — Where personal data lives
The question "can you read passwords?" has an older sibling almost nobody asks: "can you read our users' personal data?". Passwords are hashed; names, emails, phone numbers and access history are not.
LoginMaster's architecture separates the two planes: users' personal data lives exclusively in the customer tenant, while the cloud operates on encrypted data and references and never holds personal data in readable form. The cloud coordinates, signs and validates; it does not know who the people are.
POST /v1/auth/verify HTTP/1.1
Host: api.loginmaster.it
X-Project-Key: pk_live_••••••••
{ "token": "lm2f.eyJhbGciOiJFZERTQS..." }
HTTP/1.1 200 OK
{
"valid": true,
"subject": "sub_9f2a7c",
"tenant": "https://tenant.example.com",
"signatures": { "tenant": "verified", "cloud": "verified" }
}The subject is an opaque reference. Who sub_9f2a7c is, the customer tenant knows; the cloud does not. It is the same property that turns the answer to a GDPR questionnaire into a technical description rather than a commitment: where there is no readable data, there is no access to restrict.
Step 6 — The eight questions to put in writing
This is the part to take into a tender or an audit. Ask for the answers in writing, because the difference between the wordings is exactly the point.
| Question | Weak answer | Strong answer |
|---|---|---|
| Which hashing function do you use, with which parameters? | 'Industry standard' | Argon2id with stated memory cost, iterations and parallelism |
| Can an administrator set a user's password? | 'Only with authorisation' | No: the function does not exist in the system |
| Is there an impersonation feature? | 'Only for support' | No, or yes with explicit user consent and an immutable log |
| Who can disable a user's second factor? | 'Tenant administrators' | Only the user, through verified channels |
| Where does personal data live in readable form? | 'In the region you choose' | Only in the customer tenant; the central service operates on encrypted data |
| Can your staff access our users' data? | 'Only for maintenance, with tracing' | It is not in a position to do so |
| How do you prove what you claim? | A generic certification | A verifiable architectural description and a contractual clause |
| What if you receive a data access request? | 'We assess case by case' | We cannot produce data we do not hold in readable form |
The matrix of what an administrator can do
| Operation | Generalist platforms | Self-hosted IdP | LoginMaster |
|---|---|---|---|
| Read the password in clear text | No | No | No |
| Access the password hash | Typically yes for the system operator | Yes: the database is yours | No: hashes are not accessible via API |
| Set a new password for a user | Typically yes | Yes | No: the function does not exist |
| Disable a user's second factor | Typically yes | Yes | No: the user only |
| Impersonate a user | Often available | Implementable | No |
| Read users' personal data | Yes, in their own data plane | Yes | Only in the customer tenant; not in the cloud |
| Change a user's sign-in email | Typically yes | Yes | No: the user only |
The middle columns describe categories, not individual products: verify them with the specific vendor using the questions above. The product-by-product comparison is in the alternatives hub.
Why this is also a GDPR question
The Regulation requires technical measures appropriate to the risk and demands documenting who accesses personal data and on what basis. An authentication vendor is a processor, and its technical ability to access data determines the real extent of the risk to be assessed.
- A vendor that can technically access the data must be assessed on the controls it applies, its authorised staff, its sub-processors and the jurisdictions involved.
- A vendor that cannot access the data reduces the assessment to an architecture check, because the access surface does not exist.
- The difference shows up in impact assessments, in enterprise customer questionnaires and in the response to an authority's access request.
The full compliance picture — GDPR, NIS2, ISO 27001 — is on compliance, while the zero-access architecture is described on zero-knowledge authentication.
A limitation stated openly
An architecture where nobody can reset a user's credentials has a cost, and hiding it would be dishonest: there is no administrator shortcut that unblocks a user in two minutes. Recovery goes through the user's verified channels and is slower than a phone call to support. That is exactly why it is secure, and the practical countermeasure is to allow registering more than one device for the second factor, so recovery stays a rare case.
Who wrote this guide
LoginMaster is the IAM platform of CDBKR S.r.l., an Italian company that builds and operates authentication infrastructure for software houses, MSPs and European enterprise organisations. The verification questions listed here are the ones the team asked itself while designing the product, which is why the system implements neither administrator reset nor impersonation. The architectural detail is on Security; to assess your own scenario: contact us.
Frequently asked questions
Yes, but the property must be verified against the architecture, not the statements. A vendor that cannot read credentials is one where no administrator-initiated password reset exists, no user impersonation exists and the second factor can only be changed by the user. In LoginMaster passwords are protected with Argon2 and a split salt between tenant and cloud, and no administrator — customer or vendor — can reset them: the function is not implemented in the system.
No. Hashing protects against database theft, a scenario where the attacker obtains stored data. It does not protect against privilege abuse by whoever runs the system, nor against social engineering aimed at the helpdesk. If an administrator can set a new password for a user, they do not need to read the hash: replacing it is enough. In most real incidents the attacker did not break a hash, they used a legitimate function of the system.
With eight questions asked in writing: which hashing function and with which parameters, whether an administrator can set a user's password, whether an impersonation feature exists, who can disable the second factor, where personal data lives in readable form, whether vendor staff can access it, how the claims are proven and what happens on a data access request from an authority. The difference between the wordings of the answers is the real outcome of the check.
Because it proves the credential perimeter can be bypassed: if an operator can assume a user's identity without their credentials, hash strength becomes irrelevant. Impersonation is not always unacceptable — in some support contexts it is a reasonable trade-off — but it must be classified for what it is, a route into accounts, and it cannot coexist with the claim that nobody can access user data.
Those where personal data does not live at the vendor in readable form. In LoginMaster's architecture users' personal data lives exclusively in the customer tenant, while the cloud operates on encrypted data and opaque references: the cloud coordinates, signs and validates tokens, but does not know who the people are. Verifying a token returns an opaque identifier such as sub_9f2a7c, not an identity. Where there is no readable data, there is no access to restrict.
The variant (Argon2id is the current recommendation because it combines side-channel and time/memory trade-off resistance), the memory cost, the number of iterations and the degree of parallelism, plus assurance that the salt is unique per credential. Memory cost is the decisive parameter: it is what makes parallel GPU attacks disproportionately expensive, and set low, Argon2 is barely better than bcrypt. A vendor that cannot state its parameters has no verifiable answer.
One, and it is concrete: there is no administrator shortcut that unblocks a user in two minutes. Whoever loses their second-factor device gets back in through their own verified channels, a slower procedure than a phone call to support. That is the same reason the architecture is secure, given that the phone call is the most exploited attack vector. The practical countermeasure is to allow registering more than one device, so recovery stays a rare case.
Very, because an authentication vendor is a processor and its technical ability to access data determines the real extent of the risk to assess. A vendor that can access the data must be evaluated on internal controls, authorised staff, sub-processors and jurisdictions. A vendor that cannot access it reduces the assessment to an architecture check. The difference shows up in impact assessments, in enterprise customer security questionnaires and in the response to an authority's access request.
Want to see LoginMaster in action?
Request a personalized demo and discover how to manage identities and access securely and compliantly.