Kantorcodes | ℏol/acc

I just spent two days at MIT with the Advanced AI Society, NANDA, enterprises, and startups. The future of decentralized AI looks promising, enriched by brilliant minds from all walks of life. I'm grateful for the opportunity to contribute to this open-source ecosystem 🧵

HCS Trace is now protected through HOL Guard 🤝

Worst detail in this Ivanti chain: after every command the attacker put the config and the patched memory back, so Ivanti's own integrity checker came back clean. Any VPN that sat on 25.1.1.2 or 22.8R2.5 needs a look at the domain controller logs, not just the appliance's.

The part that will bite people: Langfuse's default compose file pins ClickHouse 25.12 and SigNoz's Docker docs show 25.5.6. Neither train gets this fix. Self-host either one? Override the ClickHouse image to a fixed build like 26.8.36.3.

If you run Claude Code inside Cursor, grep ~/.ssh/authorized_keys for svc-telemetry before you do anything else. Removing the extension leaves the key and the startup job behind. Linux dev boxes that already run sshd are the ones I'd check first.

The detail in Cisco's advisory most teams will miss: turning off the VPN group list does nothing if DefaultWEBVPNGroup still has a secondary auth server. Check that group first. It crashes the firewall, it does not hand it over, but it can keep crashing it.

The line in Alpine's write-up every self-hosted GitLab admin should read: they upgraded to 19.3.6 and were still exposed, because GitLab's fix ships as a patch to its bundled Ruby. If you built GitLab on your own Ruby, check ruby -v, not just the GitLab version.

HOL x AI: Episode 10 nitter.gq/i/broadcasts/1pJdRDqyZ…

The part that will bite people: 1.13.0 turns auto-login off, so a 1.12 box upgraded in place won't start until you set LANGFLOW_SUPERUSER_PASSWORD. If it ran on auto-login, the old langflow account keeps its generated password. Read the recovery steps first.

This is pretty crazy. hol.org/plugins has processed ~6m+ requests today.

A 2.7 CVSS bug on KEV is the tell here. It needs admin, so it doesn't get anyone in. What it bought WARP PANDA was ESXi commands with no log line. If your hosts don't forward syslog, fix that today along with the patch.

The detail I'd act on: Fragnesia's change lives in memory, not on disk. Reboot first and the backdoor and the evidence vanish together. Copy /etc/pam.d/su and compare a normal read with dd iflag=direct before you restart anything.

The part I'd act on: the old upgrader is what unpacks the 7.1.4 package, so today's update runs the vulnerable folder logic one last time. Add the deny rule for wp-content/upgrade/ in your server config before you click Update.

The 1.104.4 fix doesn't tighten the path check. It stops tailscaled doing socket forwarding as root, which is how OpenSSH has always worked. If you lean on Tailscale SSH check mode for prod, it doesn't cover this. Forwarding runs inside the approved session.

Adobe's July bulletin for this bug still says it isn't aware of any exploits. Sansec watched 471 Magento stores get hit in one hour back in March. Patch, then go look for the files.

The release notes for this fix say "No action needed to upgrade." If your LiteLLM proxy has an Enterprise license, ignore that line. Upgrade tonight and rotate the master key.

The first in-the-wild Brotli exploit failed: ELLIO's replay crashed Node 24.22.0 fifty times without a shell. The working chain is public, though. Run node -v inside every image you deploy today.