PKI for FortiGate, Part 6: Fail-Open, Fail-Closed and the Settings That Decide

Part 5 showed revocation working when the lists are present and current. Real networks spend time in the other states: a CRL that was never fetched, a CRL host that is down, a list that has passed its expiry date. What the FortiGate does then is controlled by three settings, and their defaults are the permissive ones.

This post tests each state and then makes a decision.

The three settings

config vpn certificate setting
    config crl-verification
        set expiry [ignore|revoke]
        set leaf-crl-absence [ignore|revoke]
        set chain-crl-absence [ignore|revoke]
    end
end
SettingControlsDefault
leaf-crl-absenceThe hub has no CRL for the CA that issued the leafignore
chain-crl-absenceThe hub has no CRL for a CA above the leaf (the intermediate)ignore
expiryA CRL the hub holds has passed its “next update” dateignore

ignore means carry on and accept. revoke means treat the problem as a revocation and reject the certificate. So ignore is fail-open and revoke is fail-closed. A fail-open setting is not a bug, but it is a decision, and the default makes it on your behalf.

Missing CRL, leaf level

I tested leaf-crl-absence by deleting the CA’s CRL object from the hub and presenting a certificate I had already revoked.

SettingWhat happened
ignoreNo CRL check ran. The revoked certificate authenticated and the tunnel came up, 1/1
revokefnbamd logged Leaf cert status is unchecked. and CRL of leaf cert is absent., validation failed, tunnel 1/0

With the default, a revoked certificate was accepted because the hub had no list to check it against. That is the worst combination: revocation silently stopped working, and the log showed a normal successful authentication.

Missing CRL, chain level

Part 5 covered this. With chain-crl-absence revoke, a missing root CRL gave CRL of cert depth 1 is absent. and a rejected authentication. I did not run the ignore case on its own. It is the default and, going by the setting’s name, would accept the chain with no CRL for the intermediate.

Expired CRL: three runs

This is the setting I understood least before testing, so I tested it three ways. I needed a CRL that had expired, which turned out to be awkward. FortiOS refuses to import a CRL that has already expired:

The upload file is expired
Command fail. Return code -542

So an expired CRL on a FortiGate can only happen the way it would in production: a CRL that was valid when loaded and then lapsed. I generated root CRLs that were valid when I loaded them and expired a few minutes later, by giving openssl explicit dates:

openssl ca -config openssl.cnf -gencrl \
    -crl_lastupdate 20261107104328Z -crl_nextupdate 20261107115128Z \
    -out root-short.crl.pem

I loaded each one, waited for the clock to pass the next-update time, and cleared the gateway. The hub and the signing workstation agreed on the time to within a few seconds.

RunCRL contentexpiryTunnelWhat the log said
1Expired, intermediate listed as revokedignoreDown***CERTIFICATE IS REVOKED*** at depth 1
2Expired, emptyignoreUpCERTIFICATE IS GOOD at both depths, no expiry warning
3Expired, emptyrevokeDownCRL has expired, then CRL is expired.

Read them in order.

Run 1: ignore does not mean the expired CRL is skipped. The list had been expired for two minutes and the hub still used it. The revocation entry was honoured and all 28 attempts were rejected. So expiry ignore ignores the staleness, not the contents.

Run 2: a stale, empty list is accepted without a word. Nineteen minutes after expiry, the hub authenticated the spoke. Both CRL checks reported GOOD. Nothing in the debug mentioned that the root CRL had expired. If your root CRL lapses and you never notice, nothing tells you.

Run 3: revoke fails closed on expiry. The same expired, empty CRL, now rejected:

[414] __check_crl-CRL has expired
[1430] fnbamd_cert_auth_copy_cert_status-CRL is expired.

All 34 attempts failed. The leaf’s own check was still GOOD, because its CRL was fresh. The expired list one level up was enough to reject it.

Taken together: with the default, an expired CRL fails open for anything not already on it, and fails closed only for entries it already contains. That is half a protection, and it fails exactly when nobody is watching.

A way to tell the failures apart

The IKE log says the same thing for every failure: certificate validation failed. The fnbamd lines differ, and so does the status value it logs at the end of the check. In my captures:

OutcomeCert st value
Authenticated290
CRL absent, or certificate revoked2a0
CRL expired12a0

I have not found these values documented, so treat them as observations from this build. The text lines above them are more reliable for diagnosis.

What fail-closed costs

Setting all three to revoke makes the whole trust model strict, and it moves the risk. The failures you accept in return:

  • A lapsed root CRL stops every tunnel at its next authentication. With a 180-day root CRL, that is a date you must write down.
  • A CA CRL that cannot be refreshed past its expiry rejects every certificate that CA issued. The CRL host becomes critical infrastructure.
  • A rebooted or reset device with no CRL loaded would, under the same logic, be unable to authenticate its peers until it has one. I did not test that case.

None of those are reasons to choose fail-open. They are reasons to monitor the CRL host, to alert on CRL days remaining, and to keep the root CRL lifetime longer than your response time. A visible outage is a better failure than a silent loss of revocation.

What I would run, and what the lab runs

The settings I would put in production:

config vpn certificate setting
    config crl-verification
        set expiry revoke
        set leaf-crl-absence revoke
        set chain-crl-absence revoke
    end
end

The lab hub ended testing with expiry revoke and chain-crl-absence revoke, and leaf-crl-absence still at ignore. That combination is inconsistent: a missing root CRL is rejected but a missing CRL for the issuing CA is accepted. I would set the third setting to revoke as well before relying on any of this.

Choosing a CRL lifetime

The root CRL changes only if an intermediate is revoked, so a long lifetime costs little. I issue it for 180 days. The only thing a short lifetime buys is a quicker lapse, which under expiry revoke is an outage. A longer lifetime does not delay a revocation either, because a revoked intermediate takes effect when a device loads the new list. How quickly the new list reaches the devices is what matters, and in this lab that is a manual paste. Part 7 covers what I would automate.