Trust

When a program on this machine validates a TLS certificate, the set of certificate authorities it will accept comes from one place: trustd, the trust store.

trustd composes that set from two inputs and serves the result.

The split is the whole design. The roots are package data — a hundred and twenty-odd certificates nobody on this machine decided anything about — and package upgrades already do the right thing with them, including removing a root Mozilla has withdrawn. The registry holds decisions: each entry is something somebody chose, which is what keeps an audit of Machine\System\Trust worth reading.

What programs actually read #

The store itself is a socket, /run/trustd/trust.sock. Peios-native programs ask it directly and are told when the answer changes, so a certificate withdrawn at ten past three stops being accepted by a running service at ten past three.

Almost nothing does that yet, so trustd also renders the store to the files the portable software ecosystem expects:

/etc/ssl/certs/ca-certificates.crt   the bundle Go and most software read
/etc/ssl/cert.pem                    OpenSSL's default CAfile
/etc/ssl/certs/<hash>.0              OpenSSL's default CApath

Those files are a copy of an answer, not the answer. Editing one changes nothing: trustd overwrites it the next time anything changes. The GenerateLinuxTrustFiles value decides whether they exist at all — see the compat files.

Seeing what is in force #

$ trust status
generation   7
health       ok
roots        121 in force
  shipped    121
  added      0
  distrusted 0
compat       GenerateLinuxTrustFiles = 1
  rendered   /system/retc/ssl/certs/ca-certificates.crt
  rendered   /system/retc/ssl/cert.pem

generation increases every time the store changes, so a program holding a copy can tell whether it is current without fetching it again. health is degraded when the last composition failed — see when the store is degraded.

Composition is all or nothing #

If the shipped bundle cannot be read, or yields implausibly few roots, or every root would be distrusted, trustd keeps whatever it rendered last and reports itself degraded. It never writes a partial store.

That is deliberate, and the reasoning is worth stating: a silently empty trust store breaks every TLS client on the machine at once, and a silently truncated one might be missing the distrust that was the point of the last change. Stale-but-whole is a recoverable state; partial is not.

A single bad entry is different. An addition that is not a certificate, is not a CA, or has expired is skipped with a warning and everything else proceeds — one administrator's typo does not take the machine's trust down.

Who may change it #

Machine\System\Trust\Certificates and its subkeys, like any registry key, are governed by their own security descriptor. By default SYSTEM and Administrators may write; everybody may read. That descriptor is the only gate: trust and reg write the same values through the same check, so there is no second permission model to keep in step with the first.

Reading is open to everyone, because the root set is public — in the default configuration it is sitting in a world-readable file.

What trustd is not #

It holds no private key. Not the machine's, not anybody's. A service that holds keys is a signing oracle if it is ever compromised, and wants almost the opposite security posture from one whose socket is open to every program on the machine. Machine identity — the certificates and keys a machine presents when it is the server — is a separate concern, and will be a separate service.

It does not fetch anything. No root is downloaded on demand and no revocation list is retrieved, in this version. What the machine trusts changes when a package is upgraded or somebody writes to the registry, and at no other time.

Two things can be decided about this machine's trust, and both are values under Machine\System\Trust\Certificates.

Trusting a certificate authority #

An internal CA, a test CA, a partner's CA:

$ trust add corp-ca /path/to/corp-root.pem
added corp-ca
  CN=Corp Issuing CA,O=Example Ltd,C=GB
  SHA-256 4f2a…c19b

The name is yours to choose — it names the decision, not the certificate, so corp-ca is a better name than 4f2a. PEM or DER is accepted; a file holding more than one certificate is refused, because a bundle added under one name would be one decision covering several authorities.

The certificate is checked before it is written: it must parse, it must be a CA (basicConstraints), and it must not have expired. trustd checks it again when it reads it — it never trusts the writer — but doing it here means a mistake is a message rather than a line in a log.

To undo:

$ trust remove corp-ca

Purposes #

By default an addition is trusted for ServerAuth — validating a TLS server. Say otherwise if you mean otherwise:

$ trust add corp-ca corp-root.pem --purposes ServerAuth,CodeSigning

This version renders ServerAuth roots and carries the rest, so a root recorded for another purpose is stored faithfully and used when the feature that consumes it arrives.

Distrusting a certificate authority #

This is the one that has to work in a hurry, and it works on any certificate — one Mozilla shipped or one somebody added:

$ trust distrust 018e13f0772532cf --reason "withdrawn, see CA/B incident 2026-08"
distrusted 018e13f0772532cf809bd1b17281867283fc48c6e13be9c69812854a490c1b05
  was: CN=DigiCert TLS ECC P384 Root G5,O=DigiCert\, Inc.,C=US

Within a second or two the root is gone from the store, gone from the rendered bundle, and its file is gone from the hashed directory. Nothing is rebuilt and nothing is rebooted, which is the reason the shipped roots can be package data in the first place: withdrawal never waits for a package.

A certificate is named by its SHA-256 fingerprint — never by subject name, which is forgeable and reused. Any of these work:

018e13f0772532cf                      the prefix `trust list` prints
018E13F0…C1B05                        upper case
01:8e:13:f0:…                          colon-separated, as other tools print it
/path/to/the-certificate.pem          the file itself

Distrusting a certificate this machine does not have is legitimate and is recorded rather than refused — the entry sits there, and if that root ever arrives in a bundle upgrade or an addition, it is already refused.

To see what has been decided, and to undo it:

$ trust list --distrusted
018e13f0772532cf  withdrawn, see CA/B incident 2026-08

1 distrusted

$ trust restore 018e13f0772532cf
restored 018e13f0772532cf809bd1b17281867283fc48c6e13be9c69812854a490c1b05

Doing it from the registry #

trust writes registry values and nothing else, so reg does the same job and is subject to exactly the same access check:

Machine\System\Trust\Certificates\
  Add\<name>\
    Certificate   REG_BINARY    the certificate, DER
    Purposes      REG_MULTI_SZ  default ServerAuth
  Distrust\
    <fingerprint> REG_SZ        why, for whoever reads it later

Write Certificate last when creating an entry by hand. Until it exists the entry is incomplete, and trustd deliberately skips incomplete entries rather than acting on half of one.

This is also how a domain distributes trust: pushing a value to Trust\Certificates\Add on every member is an ordinary policy write, not a special mechanism.

Packages may offer trust, never take it #

A package must never widen what the whole machine trusts by being installed — peipkg install quietly adding a CA is a supply-chain hole, and every system that allowed it eventually closed it.

A package that comes with a CA ships a seed (/usr/share/regim/…), inert until an image names it in [registry] autoapply or an administrator applies it. Installing the package grants nothing.

A package shipping a trust store for its own private use — a browser, a language runtime — is unaffected by any of this. Those are ordinary files that only that program reads.

When the store is degraded #

$ trust status
health       degraded — cannot read /usr/share/ca-certificates/mozilla.crt: …

degraded means the last attempt to compose the store failed, and what is rendered is what was rendered before — possibly stale, never partial. The machine keeps working with the trust it last had.

The usual causes are a missing or damaged ca-certificates package, or a distrust list that would empty the store entirely. Fix the cause and:

$ trust reload

Skipped entries are different and are not a degraded state: trust status counts them and the log names each one.

trust is the trust store's operator command. It reads over trustd's socket and writes to the registry, and the asymmetry is worth understanding: only trustd knows the effective set, because the shipped roots are package data rather than registry entries, so a listing read from the registry would show a handful of local decisions and none of the hundred and fifty roots actually in force. Writes go the other way, to the registry, so that trust and reg pass through the same access check.

Verbs #

Listing #

$ trust list
018e13f0772532cf  shipped    CN=DigiCert TLS ECC P384 Root G5,O=DigiCert\, Inc.,C=US
0a81ec5a929777f1  shipped    CN=GlobalSign,O=GlobalSign,OU=GlobalSign Root CA - R3
4f2ab19c8e10dd77  added:corp-ca  CN=Corp Issuing CA,O=Example Ltd,C=GB

121 root(s)

The first column is the first sixteen characters of the SHA-256 fingerprint, and every verb that takes a fingerprint accepts that prefix — what is printed can be pasted back in. An ambiguous prefix is refused rather than guessed at.

--purpose ServerAuth narrows the listing to roots carrying that purpose.

Showing one root #

$ trust show 4f2ab19c
subject      CN=Corp Issuing CA,O=Example Ltd,C=GB
fingerprint  4f2ab19c8e10dd77…
source       added
added as     corp-ca
purposes     ServerAuth
expires      2074003199

-----BEGIN CERTIFICATE-----
…

--pem prints the certificate alone, for piping into a file or another tool.

Exit statuses #

The separation of 2 from 1 is what lets a script tell "this machine does not trust that CA" from "I could not find out".

When a write is refused #

$ trust distrust 018e13f0772532cf
trust: not permitted to change Machine\System\Trust\Certificates\Distrust —
       changing what this machine trusts is governed by that key's descriptor

trust has no privilege of its own to grant. Changing the machine's trust is a registry write, and whether you may make it is decided by the key's security descriptor — the same answer reg would get.

The machine's trust store is a socket. The files under /etc/ssl are a compatibility rendering of it, for software that reads trust from a path — which today is very nearly all software.

They are generated. Editing one changes nothing except until the next time anything changes, and a header at the top of the bundle says so.

Three artifacts, because the ecosystem does not agree #

The hashed directory is not decoration. OpenSSL looks a certificate up there by a hash of its subject, and a stale file left behind for a root that has been distrusted would go on being trusted through that path. So the directory is replaced atomically — the new one is built alongside and exchanged in a single operation — and there is no moment at which it holds a mixture of the old set and the new.

Turning them off #

Machine\System\Trust GenerateLinuxTrustFiles decides whether the files exist:

0 is the machine whose entire software set asks the socket — an appliance image, in practice, where you know what is installed. On a general-purpose system it breaks curl, Python, Go binaries and anything else that reads a bundle, and the breakage looks like a certificate problem rather than a configuration one, so set it deliberately.

Switching to 0 removes the rendered files rather than leaving them. Stale trust left in force is worse than none: a root distrusted after the switch would go on being accepted by everything reading the abandoned copy.

$ reg set Machine/System/Trust GenerateLinuxTrustFiles dword:0
$ trust status
compat       GenerateLinuxTrustFiles = 0 (no files; the socket is the only store)

The socket keeps serving throughout — trust list still answers, and so does any program that speaks to trustd directly.

Why a copy, when name resolution gets a pointer #

/etc/resolv.conf on Peios is a constant file naming the resolver, and nothing is ever generated into it. Trust cannot work that way: there is no "ask over here" a certificate bundle can express. A file that lists authorities has to be the list.

That is the whole difference between the two, and the reason this one carries a knob and a warning while the other does not.