Pi-hole 6.9: one regex filter stops loading on upgrade

FTL 6.9 refuses one kind of regex filter at load, and the compile counter still reports it. I upgraded a 1,500-filter container to check, and the debug log leak is closed.

Pi-hole 6.9: one regex filter stops loading on upgrade

My test Pi-hole container had 1,500 regex deny filters and one odd one. I moved it to FTL v6.9, and the log said it compiled 1,501.

One of them never loaded.

Nothing turned red. DNS kept answering. The only clue was a single WARNING line, and the names that filter used to block went out to the upstream resolver. If you keep a long regex list, that's the part of this release you should check first.

How I tested: two throwaway containers on the same box, 8 vCPU, 23 GB RAM, Debian. Both images came from Docker Hub, and I've got no affiliation with the Pi-hole project. Two builds were in play. pihole/pihole:2026.10.1 ships Core v6.5, Web v6.6.1 and FTL v6.8. pihole/pihole:2026.11.0 ships Core v6.6, Web v6.6.1 and FTL v6.9.

What the 11 November release actually fixes

Pi-hole FTL v6.9 and Core v6.6 were published on 11 November 2026, and the release announcement is on the Pi-hole blog. There's no new web interface this time, so it stays at v6.6.1. This is a security release, and the FTL changelog lists every advisory.

Highest on the list is GHSA-pj9m-9623-2g75 at CVSS 7.2. A quote in the files.gravity path broke out of the SQL string in the Teleporter ATTACH statement. Together with Core's gravity.sh, that let an admin run commands as the pihole user. It needs admin access to start with, and closing it takes both halves of the update. Core v6.6 escapes the path. FTL v6.9 binds it to a prepared statement.

Then come three Core advisories at CVSS 7.3. They all involve the installer and the admin scripts following symlinks in /etc/pihole as root.

And then there's one Medium at CVSS 5.3. That's the one I'd act on today.

The debug log leaked the password hash

You can see this one for yourself, and the debug log advisory spells it out. I set the password "testpass" through FTLCONF_webserver_api_password in both containers and ran pihole -d -a in each. Line 1807 of /var/log/pihole/pihole_debug.log differed between them.

# 2026.10.1 container
pwhash = "$BALLOON-SHA256$v=1$s=1024,t=32$absXsQVfoKekbmOnAYaa8g==$IdDo4ogNHaHN1QPVmIbOYFz2zXEPTAmsSEWci2PkXig="

# 2026.11.0 container
pwhash = "<removed>"

So a debug log uploaded to tricorder before Core v6.6 carried the password hash and the TOTP secret. If you ever uploaded one, change the password and the app password, then set up 2FA again.

The regex filter that stops loading

Here's the filter that dropped out: ^(a)\1*\.tracker\.example$. The \1 is a back-reference, which means "match the exact text the first bracketed group matched, again". The * after it says that repeat can happen any number of times.

FTL v6.8 matched and blocked with it. FTL v6.9 refuses it at load with this:

Invalid regex deny filter "^(a)\1*\.tracker\.example$": Back-reference inside an unbounded repetition is not supported

Finding affected filters takes two commands. I ran this block inside the upgraded container. The first line makes FTL load every filter and keeps only the rejections. The second asks gravity.db for regex deny rows that contain \1. Mine also showed up as a WARNING in /var/log/pihole/FTL.log, which is where I found it first.

pihole-FTL regex-test nothing.invalid 2>&1 | grep -i "invalid regex"
pihole-FTL sqlite3 /etc/pihole/gravity.db \
  "SELECT id, domain FROM domainlist WHERE type=3 AND domain LIKE '%\1%';"

Here's the real output I got:

Invalid regex deny filter "^(a)\1*\.tracker\.example$": Back-reference inside an unbounded repetition is not supported
1504|^(a)\1*\.tracker\.example$

What happened on the wire

For the DNS test I kept three regex deny filters in the database and queried both versions with dig. On v6.8, a.tracker.example and aa.tracker.example both came back 0.0.0.0, which means blocked. On v6.9 the same two queries came back with no answer at all. My filter was gone, so both names went upstream and died there, which is exactly what will happen to yours if it's a real tracker.

My other two filters kept working on both versions. ads.doubleclick.net answered 0.0.0.0, and a real domain resolved normally.

The log line that lies

With 1,500 filters compiled and one rejected, FTL.log printed this:

Compiled 0 allow and 1501 deny regex for 3 clients in 146.0 msec

That 1501 matches the number of rows in the database. It doesn't count the filters that work.

Why does that matter? It's the line I'd glance at after an upgrade to confirm my lists came through, and it says yes. There's no failing health check and no red banner. DNS answers queries normally, so from the outside the container looks healthy. One WARNING line is the only sign of trouble, and the domains your dropped filter used to catch now reach your upstream as if you'd never written the rule. A counter that counts rows and calls it compiling is an optimistic counter.

Speed with 1,500 filters

Speed is where this release pays you back. With 1,500 stored regex deny filters, I took the median of 7 runs of pihole-FTL regex-test nothing-matches-here.example.org on each version, same host.

  • FTL v6.8: full scan of all 1,500 filters in 0.879 ms, filter compile 171.05 ms
  • FTL v6.9: full scan in 0.134 ms, filter compile 128.52 ms

That's about 6.6x faster to scan and 1.3x faster to compile. A domain that does match took 0.909 ms on 6.8 and 0.122 ms on 6.9.

How to upgrade

Back up first with Teleporter. It's under Settings -> Teleporter in the web interface, or run pihole-FTL --teleporter.

On Docker, the image tag is 2026.11.0. Pull it and recreate the container:

docker compose pull && docker compose up -d

Don't try pihole -up inside the container. It answers "Function not supported in Docker images". On bare metal, pihole -up is the right command.

I also tested an in-place upgrade. I copied /etc/pihole from the 6.8 container into a container running the 2026.11.0 image. All 1,500 filters were still in gravity.db afterwards. FTL reported v6.9, Core v6.6 and Web v6.6.1, and DNS answered queries normally. That bad filter was dropped with the WARNING line, and the other 1,500 kept blocking.

What I couldn't reproduce, and what I didn't test

The release notes say an omitted lower bound in a regex, {,n}, now means {0,n}. I tested a{,3} against zero to three a's and both versions behaved identically. {,} was identical too. So I saw no user-visible change from that line.

My tests couldn't confirm the "more than three nested approximate-matching groups" rejection. Patterns using approximate groups returned NO MATCH on both versions, so I've got no evidence either way.

Scope, plainly: amd64 only. I ran no DNSSEC or DHCP tests, and I didn't attempt to exploit the files.gravity issue.

A filter with a back-reference inside an unbounded repetition won't survive FTL v6.9, and the compile count won't warn you. Run the check before you pull the image.