A binary runs out of /tmp. A second later it attaches a debugger to a
system service and starts sweeping the subnet. On most boxes that's three
log lines in three places, and you read them on Monday. Aegis reads it as
one story. It watches every process tree from the kernel with eBPF, adds
the signals up into a single score, and opens one incident with the
evidence attached. Switch on enforce mode and it acts: the tree is held
in the kernel, frozen and cut off from the network, alive and waiting
for you to decide.
I built it for my own Rocky Linux servers. It's Rust, it ships as an RPM with its own SELinux policy, and it starts out only watching.
That's a real run of examples/lab/run.sh on a Rocky
10.2 VM. It uses copies of the host's own bash and strace, never leaves
the box, and cleans up after itself, so you can run it on a test VM and
watch the same thing happen. This run is in enforce mode; in audit mode
you get the same incident, and nothing is touched.
Rocky Linux 10, x86_64 or aarch64. Every release is signed and ships with SBOMs and build provenance; Installation shows how to check one before you install it.
VERSION=0.2.3 ARCH=x86_64 # or aarch64
base=https://github.com/jamesconwaydev/aegis-v2/releases/download/v$VERSION
curl -fsSLO "$base/aegis-v2-$VERSION-1.el10.$ARCH.rpm" \
-O "$base/aegis-v2-selinux-$VERSION-1.el10.noarch.rpm"
dnf install "./aegis-v2-$VERSION-1.el10.$ARCH.rpm" \
"./aegis-v2-selinux-$VERSION-1.el10.noarch.rpm"
aegisctl preflight
systemctl enable --now aegisdpreflight checks for kernel BTF, cgroup v2 and the BPF LSM, and tells
you the fix for anything missing. Rocky 10.2 has all three out of the box.
Leave it in audit mode for two or three days of normal work. It scores and records everything and touches nothing. Then look at what it found:
aegisctl incidents list
aegisctl incidents show <id>Each incident is either worth your time or a false positive you
allowlist. Once the list goes quiet, set mode = "enforce" in
/etc/aegis-v2/config.toml and systemctl reload aegisd. From then on,
a held tree waits for one of two commands: aegisctl incidents release <id> gives it back, and aegisctl incidents kill <id> ends it.
- Watches from the kernel. eBPF probes on exec (the real path, not
argv[0]), fork, exit, ptrace attach,memfd_createand outbound TCP connects. Each new program's environment is read at exec, before its first instruction, so a preloaded library can't scrub itself before the check. - Scores behaviour, not names. An exec from
/tmpadds 25, a memfd exec 65, an attach to a system daemon 45, a connect burst 40. Scores halve every 30 seconds of quiet, and a whole tree counts as one. - Trusts binaries you can prove.
dnf,gdbandpodmanonly get a pass when the file on disk is RPM-owned and unmodified, or matches a hash you pinned, with noLD_PRELOADor debugger in sight. - Writes incidents, not alerts. One directory per incident: the tree, the evidence, and a timeline of every decision. Out to journald, JSON-lines, OTLP logs and metrics, and webhooks (JSON or Slack).
- Throttles at 50, quarantines at 80, in enforce mode. Throttle caps the tree's CPU and memory in its own scope. Quarantine holds it with BPF LSM programs (no fork, connect, mount, ptrace, module loads or writes), freezes it with systemd, and drops its traffic in nftables. Nothing gets killed unless you say so.
- Runs tight. Six capabilities instead of all of them, a read-only
filesystem outside its own state, socket and log directories, its own
SELinux domain, and a control socket that answers root only.
/etcis left byte for byte as it was.
Better you hear it here than find out during an incident.
- Root who knows it's there. Root can stop aegisd and unload its probes. Ship telemetry off the box so the record survives.
- UDP, and scans split across many processes. Only TCP connects count, per process.
sh /tmp/x. The kernel ran the shell, so that's what Aegis sees. A script run directly through its#!line is caught.- Writes made before the hold. The hold lands milliseconds after the score crosses 80, and anything written before then is on disk.
- The seconds after an nftables reload.
systemctl reload nftablesflushes every quarantined tree's drop rule. aegisd puts them back within about 10 seconds. Until then the tree is still frozen and held, but nothing drops its traffic.
The threat model lists every known gap and why it's there.
Everything lives in the wiki: installation and uninstalling,
configuration, scoring and baselining, reading incidents, telemetry,
aegisctl, SELinux, the threat model, troubleshooting and the
roadmap. The same pages are in docs/,
versioned with each release.
False positives are the most useful thing you can send right now: there's an issue form for them. Questions and ideas go in Discussions, and the good first issues each name the file to change and the test to add. Read CONTRIBUTING.md before you send code.
Found a security problem? Report it privately; see SECURITY.md.
aegisd, aegisctl and their libraries are dual-licensed under
MIT or Apache-2.0, at your option. The
eBPF programs in aegis-ebpf are GPL-2.0-only,
because the kernel requires it.
