Add audit/ with 8 named CVE regression checks against doos
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
This commit is contained in:
Claude Sonnet 5.5
@@ -48,6 +48,7 @@ Never hand-edit `VERSION` or create tags manually. Never skip the tests to relea
|
||||
| These workflow rules | this file |
|
||||
| A measurement, or anything the charts show | `make bench` then `make charts` (rewrites `docs/charts/*.svg` and every `<!--bench:...-->` number and table in `COMPARISON.md` and `docs/RESEARCH.md`); never type those numbers by hand |
|
||||
| A feature or limit that appears in the comparison tables | `COMPARISON.md` (tables, the gaps list, the conclusion) |
|
||||
| A code path one of the `audit/CVE-*` checks targets (overlay writes, tar extraction, mount/directory creation, net namespace sharing, seccomp denylist) | re-run `make audit`; update that CVE's `README.md` if the verdict's reasoning changed |
|
||||
| A new or renamed markdown file | `scripts/check-docs.sh` DOCS list, the nav line (`**Docs:**`) of every document, `README.md` Project docs |
|
||||
| Image catalog, sources, verification, pinned keys | `SPEC.md` Images, `README.md`, `docs/RESEARCH.md` |
|
||||
| Seccomp list, capability set, `/etc` allowlist | `SPEC.md` Hardening and filesystem, `README.md`, `docs/RESEARCH.md` |
|
||||
@@ -84,9 +85,10 @@ installed once for the binary being tested (`doos apparmor`; needs sudo, ask the
|
||||
|
||||
## Make targets
|
||||
|
||||
`make help` lists them: `build`, `test`, `check-docs`, `bench` (re-measure into `docs/charts/bench.json`),
|
||||
`make help` lists them: `build`, `test`, `audit` (named CVE regression checks against `build/doos`,
|
||||
see [audit/README.md](audit/README.md)), `check-docs`, `bench` (re-measure into `docs/charts/bench.json`),
|
||||
`charts` (SVGs and generated numbers from that file), `docs`, `release KIND=patch MSG="..."`, `clean`.
|
||||
The Makefile is a thin wrapper; the logic lives in `CMakeLists.txt`, `scripts/` and `docs/charts/`.
|
||||
The Makefile is a thin wrapper; the logic lives in `CMakeLists.txt`, `scripts/`, `audit/` and `docs/charts/`.
|
||||
|
||||
## Environment notes
|
||||
|
||||
|
||||
@@ -2,6 +2,11 @@
|
||||
|
||||
**Docs:** [README](README.md) · [SPEC](SPEC.md) · [ARCHITECTURE](docs/ARCHITECTURE.md) · [RESEARCH](docs/RESEARCH.md) · [COMPARISON](COMPARISON.md) · [AGENTS](AGENTS.md) · [CHANGELOG](CHANGELOG.md)
|
||||
|
||||
## 1.1.5 - 2026-10-23
|
||||
- `audit/`: 8 named, sourced CVEs relevant to doos's own mechanisms (overlay writes, tar extraction, mount/directory creation, net namespace sharing, the mount-family seccomp denylist, kernel user-namespace exposure), each with its own `README.md`, `Makefile` and non-destructive `reproduce.sh` against `build/doos`. `make audit` runs them concurrently (`-j8`, recursive make); `make -C audit list|report` also available. Kernel memory-corruption CVEs (Dirty Pipe, Dirty COW, `legacy_parse_param`) are checked by kernel version only, no exploit code.
|
||||
- Found by the audit and disclosed: `fsutil::mkdirs` follows a symlink planted at a container's path, so a symlink under `$DOOS_HOME/containers/<name>` redirects where that container's state is written (low severity: needs write access to the caller's own `$DOOS_HOME`). Tracked as gap G8 in [RESEARCH](docs/RESEARCH.md) and in [COMPARISON](COMPARISON.md); not yet fixed.
|
||||
- Docs: `scripts/check-docs.sh` excludes `audit/*.md` (a separate, self-contained doc set); `README.md` and `AGENTS.md` link to `audit/README.md`.
|
||||
|
||||
## 1.1.4 - 2026-10-23
|
||||
- Docs: [COMPARISON](COMPARISON.md) section 6.1 answers whether doos is a security-wise Docker replacement (same class as rootless Podman, not for untrusted code), summarized in the README safety summary.
|
||||
|
||||
|
||||
+2
-1
@@ -166,7 +166,7 @@ Short answer: it is in the same class as rootless Podman and rootless Docker by
|
||||
|
||||
Use doos for your own tools, builds and experiments, where you accept the risk. For code you do not trust, or for several tenants, use rootless Podman with its default seccomp and SELinux or AppArmor, and gVisor or Kata when the kernel must not be shared (see the conclusion table in section 8). Functionally it is also not a drop-in replacement: see section 7.
|
||||
|
||||
Full list of what doos does not cover is at the end of [RESEARCH](docs/RESEARCH.md); the exact hardening is in [SPEC](SPEC.md) under Hardening.
|
||||
Full list of what doos does not cover is at the end of [RESEARCH](docs/RESEARCH.md); the exact hardening is in [SPEC](SPEC.md) under Hardening. Named CVE regression checks against `build/doos` live in [audit/README.md](audit/README.md) (`make audit`).
|
||||
|
||||
## 7. What doos lacks, and which code would change
|
||||
|
||||
@@ -183,6 +183,7 @@ These are the gaps that matter when comparing, each next to where it lives. They
|
||||
| Linux x86_64 tested only | `src/hardening.cpp` guards aarch64, never run there |
|
||||
| seccomp is a denylist | `deniedSyscalls` in `src/hardening.cpp` |
|
||||
| Test coverage: one machine | `tests/run.sh` |
|
||||
| `fsutil::mkdirs` follows a planted symlink at a container's path, so a symlink at `$DOOS_HOME/containers/<name>` redirects where that container's state is written (low severity on a single-user `$DOOS_HOME`; found by [`audit/CVE-2021-30465`](audit/CVE-2021-30465/README.md)) | `fsutil::mkdirs` in `src/util.cpp` |
|
||||
|
||||
## 8. Conclusion
|
||||
|
||||
|
||||
@@ -4,7 +4,7 @@
|
||||
BUILD ?= build
|
||||
DOOS := $(BUILD)/doos
|
||||
|
||||
.PHONY: help build test check-docs bench charts docs release clean
|
||||
.PHONY: help build test audit check-docs bench charts docs release clean
|
||||
|
||||
help: ## list the targets
|
||||
@grep -hE '^[a-z-]+:.*##' $(MAKEFILE_LIST) | sed -E 's/:.*## /\t/' | column -t -s "$$(printf '\t')"
|
||||
@@ -16,6 +16,9 @@ build: ## configure and compile into build/
|
||||
test: build ## end-to-end suite (real downloads, needs network, pasta, gpg, cgroup v2)
|
||||
tests/run.sh $(DOOS)
|
||||
|
||||
audit: build ## run the CVE regression checks under audit/ against build/doos (concurrent, -j)
|
||||
$(MAKE) -C audit -j8 test DOOS=$(abspath $(DOOS))
|
||||
|
||||
check-docs: ## every document links to every other, every relative link resolves
|
||||
scripts/check-docs.sh
|
||||
|
||||
|
||||
@@ -187,5 +187,6 @@ Details: [COMPARISON](COMPARISON.md) section 6.1.
|
||||
| [`COMPARISON.md`](COMPARISON.md) | doos next to Docker, Podman, bubblewrap, nspawn, LXC/Incus, Apptainer, Distrobox, gVisor/Kata, with pros, cons, measurements and a conclusion |
|
||||
| [`AGENTS.md`](AGENTS.md) | Rules for AI agents, release workflow |
|
||||
| [`CHANGELOG.md`](CHANGELOG.md) | Release history |
|
||||
| [`audit/README.md`](audit/README.md) | Named CVE regression checks run against `build/doos` (`make audit`) |
|
||||
|
||||
Not sure whether doos is the right tool? Read the conclusion of [COMPARISON](COMPARISON.md) first.
|
||||
|
||||
@@ -0,0 +1,7 @@
|
||||
# CVE-2016-5195: see README.md in this directory.
|
||||
DOOS ?= $(abspath ../../build/doos)
|
||||
export DOOS
|
||||
|
||||
.PHONY: test
|
||||
test:
|
||||
@./reproduce.sh "$(DOOS)"
|
||||
@@ -0,0 +1,35 @@
|
||||
# CVE-2016-5195: "Dirty COW": race condition in copy-on-write handling of private read-only mappings
|
||||
|
||||
**Scope:** [audit](../README.md) · [README](../../README.md) · [COMPARISON](../../COMPARISON.md) · [docs/RESEARCH](../../docs/RESEARCH.md)
|
||||
|
||||
CVSS 7.0-7.2. Fixed in Linux 4.8.3 (and backported to every maintained branch since).
|
||||
|
||||
## Source
|
||||
|
||||
- [Ubuntu advisory](https://ubuntu.com/security/CVE-2016-5195)
|
||||
- [NVD entry](https://nvd.nist.gov/vuln/detail/CVE-2016-5195)
|
||||
|
||||
## Relevance to doos
|
||||
|
||||
A race in the memory manager's copy-on-write handling let a local user write to memory
|
||||
that should have been read-only, including on-disk binaries mapped read-only. Like
|
||||
Dirty Pipe, this needs no container-specific trigger; any unprivileged process on a
|
||||
vulnerable kernel is affected, with or without doos.
|
||||
|
||||
|
||||
## What the check does
|
||||
|
||||
Kernel version only. Fixed in 4.8.3 and every branch since; on a host running anything
|
||||
from the last several years this is expected to PASS trivially. It is kept in the list
|
||||
to complete the "shared-kernel" risk category doos itself names in
|
||||
COMPARISON.md section 6, not because it is expected to ever FAIL on a maintained host.
|
||||
|
||||
|
||||
## Caveats
|
||||
|
||||
Same version-check limits as CVE-2022-0847: distribution kernels sometimes backport
|
||||
fixes without a version bump this check recognizes; it reports the implication of the
|
||||
reported version, not a scan of the kernel's actual code.
|
||||
|
||||
|
||||
Run: `make test` in this directory, or `make -C audit test` from the repo root.
|
||||
Executable
+25
@@ -0,0 +1,25 @@
|
||||
#!/usr/bin/env bash
|
||||
# CVE-2016-5195 (Dirty COW): kernel version against the fixed release. No exploit code.
|
||||
set -u
|
||||
skip() { echo "SKIP: $*"; exit 0; }
|
||||
fail() { echo "FAIL: $*"; exit 1; }
|
||||
pass() { echo "PASS: $*"; exit 0; }
|
||||
|
||||
ver=$(uname -r | grep -oE '^[0-9]+\.[0-9]+(\.[0-9]+)?')
|
||||
major=$(echo "$ver" | cut -d. -f1)
|
||||
minor=$(echo "$ver" | cut -d. -f2)
|
||||
patch=$(echo "$ver" | cut -d. -f3); patch=${patch:-0}
|
||||
|
||||
if [ "$major" -gt 4 ] 2>/dev/null; then
|
||||
pass "kernel $ver, major version past 4.x, long past the Dirty COW fix (4.8.3)"
|
||||
fi
|
||||
if [ "$major" -eq 4 ] && [ "$minor" -gt 8 ]; then
|
||||
pass "kernel $ver is past 4.8, past the Dirty COW fix"
|
||||
fi
|
||||
if [ "$major" -eq 4 ] && [ "$minor" -eq 8 ] && [ "$patch" -ge 3 ]; then
|
||||
pass "kernel $ver is at or past the fixed 4.8.3"
|
||||
fi
|
||||
if [ "$major" -lt 2 ] 2>/dev/null; then
|
||||
skip "could not parse kernel version from 'uname -r': $(uname -r)"
|
||||
fi
|
||||
fail "kernel $ver predates the Dirty COW fix (4.8.3); this bug has existed since 2.6.22"
|
||||
@@ -0,0 +1,7 @@
|
||||
# CVE-2019-5736: see README.md in this directory.
|
||||
DOOS ?= $(abspath ../../build/doos)
|
||||
export DOOS
|
||||
|
||||
.PHONY: test
|
||||
test:
|
||||
@./reproduce.sh "$(DOOS)"
|
||||
@@ -0,0 +1,43 @@
|
||||
# CVE-2019-5736: runc container escape via /proc/self/exe overwrite
|
||||
|
||||
**Scope:** [audit](../README.md) · [README](../../README.md) · [COMPARISON](../../COMPARISON.md) · [docs/RESEARCH](../../docs/RESEARCH.md)
|
||||
|
||||
CVSS 7.2. Fixed in runc 1.0.0-rc7, Docker 18.09.2.
|
||||
|
||||
## Source
|
||||
|
||||
- [Ubuntu security notice](https://discourse.ubuntu.com/t/runc/55877)
|
||||
- [Trend Micro writeup](https://www.trendmicro.com/vinfo/be/security/news/vulnerabilities-and-exploits/cve-2019-5736-runc-container-escape-vulnerability-provides-root-access-to-the-target-machine)
|
||||
|
||||
## Relevance to doos
|
||||
|
||||
runc re-executes itself through `/proc/self/exe` as part of its container-start
|
||||
sequence. A root process inside the container could replace its own `/bin/sh` with a
|
||||
script whose interpreter line was `#!/proc/self/exe`, so the *host's* runc binary ran
|
||||
in place of the container's shell, then race a write to that open file descriptor to
|
||||
overwrite the host runc binary itself.
|
||||
|
||||
doos has no equivalent re-exec step: `doos attach` builds the namespace set and calls
|
||||
the target command directly from the same process, there is no second doos invocation
|
||||
from inside the container. The applicable half of this bug class for doos is the other
|
||||
one: in host mode the container's `/usr` is the *same files* as the host's, through
|
||||
overlayfs, so a write to a binary the container can see (such as doos's own binary, if
|
||||
installed under `/usr/local/bin`) must never reach the host's real file.
|
||||
|
||||
|
||||
## What the check does
|
||||
|
||||
Builds a host-mode container, overwrites a file under the overlaid `/usr` from inside
|
||||
it (anything, not specifically doos's own binary, to keep the check independent of
|
||||
where this `doos` happens to be installed), exits, then compares the real host file's
|
||||
checksum before and after. A difference is FAIL.
|
||||
|
||||
|
||||
## Caveats
|
||||
|
||||
This does not reproduce the runc bug's mechanism (there is no self re-exec to race);
|
||||
it tests the property that would have mattered if it did: host binaries are read-only
|
||||
from a container's point of view. Needs `sha256sum` and a host-mode attach to work.
|
||||
|
||||
|
||||
Run: `make test` in this directory, or `make -C audit test` from the repo root.
|
||||
Executable
+31
@@ -0,0 +1,31 @@
|
||||
#!/usr/bin/env bash
|
||||
# CVE-2019-5736: host binaries must stay unwritable from inside a host-mode container.
|
||||
set -u
|
||||
DOOS="${1:-}"
|
||||
name() { echo "$0"; }
|
||||
skip() { echo "SKIP: $*"; exit 0; }
|
||||
fail() { echo "FAIL: $*"; exit 1; }
|
||||
pass() { echo "PASS: $*"; exit 0; }
|
||||
|
||||
[ -x "$DOOS" ] || skip "no doos binary at '$DOOS'"
|
||||
command -v sha256sum >/dev/null || skip "sha256sum not installed"
|
||||
|
||||
TARGET=/usr/bin/true
|
||||
[ -r "$TARGET" ] || skip "$TARGET not present or not readable on this host"
|
||||
|
||||
TMP=$(mktemp -d)
|
||||
trap 'rm -rf "$TMP"' EXIT
|
||||
export DOOS_HOME="$TMP/home"
|
||||
|
||||
before=$(sha256sum "$TARGET" | cut -d' ' -f1)
|
||||
|
||||
out=$("$DOOS" attach --net none audit5736 -- sh -c "echo clobber > '$TARGET' 2>/dev/null; echo done" 2>&1)
|
||||
rc=$?
|
||||
"$DOOS" delete --yes audit5736 >/dev/null 2>&1
|
||||
|
||||
after=$(sha256sum "$TARGET" | cut -d' ' -f1)
|
||||
|
||||
if [ "$before" != "$after" ]; then
|
||||
fail "host $TARGET changed after a container wrote to the overlaid path (before=$before after=$after)"
|
||||
fi
|
||||
pass "host $TARGET unchanged after an attempted overwrite from inside the container (doos exit=$rc)"
|
||||
@@ -0,0 +1,7 @@
|
||||
# CVE-2020-15257: see README.md in this directory.
|
||||
DOOS ?= $(abspath ../../build/doos)
|
||||
export DOOS
|
||||
|
||||
.PHONY: test
|
||||
test:
|
||||
@./reproduce.sh "$(DOOS)"
|
||||
@@ -0,0 +1,42 @@
|
||||
# CVE-2020-15257: containerd-shim API reachable from a container sharing the host network namespace
|
||||
|
||||
**Scope:** [audit](../README.md) · [README](../../README.md) · [COMPARISON](../../COMPARISON.md) · [docs/RESEARCH](../../docs/RESEARCH.md)
|
||||
|
||||
CVSS 5.2 (Medium). Fixed in containerd 1.3.9 / 1.4.3.
|
||||
|
||||
## Source
|
||||
|
||||
- [Ubuntu advisory](https://ubuntu.com/security/CVE-2020-15257)
|
||||
- [Tenable entry](https://www.tenable.com/cve/CVE-2020-15257)
|
||||
|
||||
## Relevance to doos
|
||||
|
||||
containerd-shim listened on an abstract Unix domain socket and only checked the caller's
|
||||
effective UID (0), not which namespace it called from. A container sharing the host's
|
||||
network namespace (abstract sockets are namespaced by the *network* namespace, not the
|
||||
mount namespace) could reach that socket and ask the shim to start processes with the
|
||||
shim's own privileges.
|
||||
|
||||
doos documents this exact mechanism as a known, explicit trade-off of `--net host`: it
|
||||
shares the host's abstract socket namespace and prints a note recommending `pasta`
|
||||
instead (README "Networking" table, COMPARISON.md section 6). This check turns that
|
||||
documentation claim into a measurement: a container must get its own network namespace
|
||||
in every mode except the explicit, opt-in `--net host`.
|
||||
|
||||
## What the check does
|
||||
|
||||
Abstract unix sockets are namespaced by the *network* namespace, not the mount namespace, so
|
||||
the mechanism this CVE needs is simply: does the container get its own network namespace at
|
||||
all. The check compares `readlink /proc/self/ns/net` from the host against the same read from
|
||||
inside a `--net none`, a default-mode, and a `--net host` container. The first two must differ
|
||||
from the host's (their own namespace, host abstract sockets out of view); `--net host` sharing
|
||||
the host's is the documented, opt-in exposure, not a regression.
|
||||
|
||||
## Caveats
|
||||
|
||||
Does not open an actual abstract socket and attempt a cross-namespace connection; it checks the
|
||||
namespace identity directly, which is what determines reachability for every process in that
|
||||
namespace, not only the one this script happens to start. `pasta` is not checked here, since the
|
||||
README already measures it as using a private socket namespace.
|
||||
|
||||
Run: `make test` in this directory, or `make -C audit test` from the repo root.
|
||||
Executable
+40
@@ -0,0 +1,40 @@
|
||||
#!/usr/bin/env bash
|
||||
# CVE-2020-15257: a container must get its own network namespace (which removes the host's
|
||||
# abstract unix sockets from view) in every net mode except the explicit, documented --net host.
|
||||
set -u
|
||||
DOOS="${1:-}"
|
||||
skip() { echo "SKIP: $*"; exit 0; }
|
||||
fail() { echo "FAIL: $*"; exit 1; }
|
||||
pass() { echo "PASS: $*"; exit 0; }
|
||||
|
||||
[ -x "$DOOS" ] || skip "no doos binary at '$DOOS'"
|
||||
|
||||
TMP=$(mktemp -d)
|
||||
trap 'rm -rf "$TMP"' EXIT
|
||||
export DOOS_HOME="$TMP/home"
|
||||
|
||||
host_ns=$(readlink /proc/self/ns/net 2>/dev/null) || skip "cannot read /proc/self/ns/net on this host"
|
||||
|
||||
none_ns=$("$DOOS" attach --net none auditns_none -- readlink /proc/self/ns/net 2>/dev/null | tail -1)
|
||||
"$DOOS" delete --yes auditns_none >/dev/null 2>&1
|
||||
|
||||
default_ns=$("$DOOS" attach auditns_default -- readlink /proc/self/ns/net 2>/dev/null | tail -1)
|
||||
"$DOOS" delete --yes auditns_default >/dev/null 2>&1
|
||||
|
||||
host_mode_ns=$("$DOOS" attach --net host auditns_host -- readlink /proc/self/ns/net 2>/dev/null | tail -1)
|
||||
"$DOOS" delete --yes auditns_host >/dev/null 2>&1
|
||||
|
||||
if [ -z "$none_ns" ] || [ -z "$default_ns" ] || [ -z "$host_mode_ns" ]; then
|
||||
skip "could not read a container's net namespace id (host=$host_ns none=$none_ns default=$default_ns host-mode=$host_mode_ns)"
|
||||
fi
|
||||
|
||||
if [ "$none_ns" = "$host_ns" ]; then
|
||||
fail "--net none shares the host's network namespace ($none_ns): host abstract sockets would be reachable"
|
||||
fi
|
||||
if [ "$default_ns" = "$host_ns" ]; then
|
||||
fail "the default net mode shares the host's network namespace ($default_ns): host abstract sockets would be reachable"
|
||||
fi
|
||||
if [ "$host_mode_ns" != "$host_ns" ]; then
|
||||
pass "--net host did not share the host's namespace ($host_mode_ns vs $host_ns); isolation is stronger than documented, default modes are isolated ($none_ns, $default_ns)"
|
||||
fi
|
||||
pass "default/none use a private net namespace ($none_ns, $default_ns), --net host shares the host's ($host_ns) as documented"
|
||||
@@ -0,0 +1,7 @@
|
||||
# CVE-2021-30465: see README.md in this directory.
|
||||
DOOS ?= $(abspath ../../build/doos)
|
||||
export DOOS
|
||||
|
||||
.PHONY: test
|
||||
test:
|
||||
@./reproduce.sh "$(DOOS)"
|
||||
@@ -0,0 +1,46 @@
|
||||
# CVE-2021-30465: runc symlink-exchange race during mount setup ("symlink-exchange attack")
|
||||
|
||||
**Scope:** [audit](../README.md) · [README](../../README.md) · [COMPARISON](../../COMPARISON.md) · [docs/RESEARCH](../../docs/RESEARCH.md)
|
||||
|
||||
CVSS 8.5. Fixed in runc 1.0.0-rc95.
|
||||
|
||||
## Source
|
||||
|
||||
- [Red Hat RHSB-2021-004](https://access.redhat.com/security/vulnerabilities/RHSB-2021-004)
|
||||
- [GitHub advisory GHSA-c3xm-pvg7-gh7r](https://github.com/opencontainers/runc/security/advisories/GHSA-c3xm-pvg7-gh7r)
|
||||
|
||||
## Relevance to doos
|
||||
|
||||
runc resolved a mount target, then mounted it, with a window between the two where an
|
||||
attacker sharing a volume could swap the target for a symlink, redirecting the mount
|
||||
outside the container's intended root.
|
||||
|
||||
doos builds its own container directories with `fsutil::mkdirs`, which calls
|
||||
`std::filesystem::create_directories`. That call follows an existing symlink at any
|
||||
path component rather than refusing it, the same general shape of bug (trusting a path
|
||||
component that could have been swapped for a symlink by something else with write
|
||||
access to the same parent directory), applied to `$DOOS_HOME/containers/<name>` instead
|
||||
of a mounted volume. The severity is much lower here: it needs write access to the
|
||||
caller's own `$DOOS_HOME`, which on a single-user system is the same trust boundary as
|
||||
running `doos` itself, not a separate tenant as in the runc case.
|
||||
|
||||
|
||||
## What the check does
|
||||
|
||||
Plants a symlink at the exact path `get_or_create` would otherwise create
|
||||
(`$DOOS_HOME/containers/<name>`), pointing at a sentinel directory outside
|
||||
`$DOOS_HOME`. Runs `doos get_or_create` for that name and checks whether any container
|
||||
state ended up written through the symlink into the sentinel directory. Presence there
|
||||
is FAIL; doos either refusing to proceed or creating state only inside `$DOOS_HOME`
|
||||
(symlink left untouched or replaced) is PASS.
|
||||
|
||||
|
||||
## Caveats
|
||||
|
||||
This is a same-user, same-`$DOOS_HOME` scenario, not a multi-tenant one: on a normal
|
||||
single-user install nothing else has write access to plant the symlink in the first
|
||||
place, which is why this is a defense-in-depth check and not a finding about a
|
||||
realistic attacker in doos's actual (single-user) deployment model.
|
||||
|
||||
|
||||
Run: `make test` in this directory, or `make -C audit test` from the repo root.
|
||||
Executable
+29
@@ -0,0 +1,29 @@
|
||||
#!/usr/bin/env bash
|
||||
# CVE-2021-30465-class: a symlink planted at a container's intended directory path must not
|
||||
# cause doos to write container state outside $DOOS_HOME.
|
||||
set -u
|
||||
DOOS="${1:-}"
|
||||
skip() { echo "SKIP: $*"; exit 0; }
|
||||
fail() { echo "FAIL: $*"; exit 1; }
|
||||
pass() { echo "PASS: $*"; exit 0; }
|
||||
|
||||
[ -x "$DOOS" ] || skip "no doos binary at '$DOOS'"
|
||||
|
||||
TMP=$(mktemp -d)
|
||||
trap 'rm -rf "$TMP"' EXIT
|
||||
export DOOS_HOME="$TMP/home"
|
||||
SENTINEL="$TMP/sentinel-outside-home"
|
||||
mkdir -p "$DOOS_HOME" "$SENTINEL"
|
||||
|
||||
ln -s "$SENTINEL" "$DOOS_HOME/containers" 2>/dev/null || {
|
||||
mkdir -p "$DOOS_HOME"
|
||||
ln -s "$SENTINEL" "$DOOS_HOME/containers"
|
||||
}
|
||||
|
||||
"$DOOS" get_or_create audit30465 >"$TMP/out.log" 2>&1
|
||||
rc=$?
|
||||
|
||||
if [ -e "$SENTINEL/audit30465" ]; then
|
||||
fail "container state was written through the planted symlink into $SENTINEL (exit=$rc)"
|
||||
fi
|
||||
pass "no container state written outside \$DOOS_HOME through the planted symlink (exit=$rc); see $TMP/out.log"
|
||||
@@ -0,0 +1,7 @@
|
||||
# CVE-2022-0185: see README.md in this directory.
|
||||
DOOS ?= $(abspath ../../build/doos)
|
||||
export DOOS
|
||||
|
||||
.PHONY: test
|
||||
test:
|
||||
@./reproduce.sh "$(DOOS)"
|
||||
@@ -0,0 +1,47 @@
|
||||
# CVE-2022-0185: kernel heap overflow in legacy_parse_param, reachable from an unprivileged user namespace
|
||||
|
||||
**Scope:** [audit](../README.md) · [README](../../README.md) · [COMPARISON](../../COMPARISON.md) · [docs/RESEARCH](../../docs/RESEARCH.md)
|
||||
|
||||
CVSS 8.4 (NVD), CISA KEV. Introduced in Linux 5.1-rc1, fixed 18 Jan 2022.
|
||||
|
||||
## Source
|
||||
|
||||
- [oss-sec disclosure](https://seclists.org/oss-sec/2022/q1/55)
|
||||
- [Hack The Box case study](https://www.hackthebox.com/blog/cve-2022-0185_a_case_study)
|
||||
|
||||
## Relevance to doos
|
||||
|
||||
An integer underflow in `legacy_parse_param` (`fs/fs_context.c`) let a filesystem mount
|
||||
option string trigger a heap buffer overflow. The reporter's PoC needed only
|
||||
`CAP_SYS_ADMIN`, which an unprivileged user can hold inside their own user namespace via
|
||||
plain `unshare`, so this was a real unprivileged-to-root path on any host with user
|
||||
namespaces enabled and an unpatched kernel, independent of which container tool, if
|
||||
any, is in use.
|
||||
|
||||
doos relies on unprivileged user namespaces for everything, so a kernel at this bug
|
||||
class is a host risk no matter what doos's own code does. The only thing doos's own
|
||||
code can do about a kernel bug is reduce the syscalls a container can reach, which is
|
||||
exactly what its seccomp denylist does for the mount family.
|
||||
|
||||
|
||||
## What the check does
|
||||
|
||||
Two independent, non-destructive checks, both required for PASS:
|
||||
1. Kernel version against the fixed table (`uname -r`). Below the fixed version for its
|
||||
major branch is FAIL (this kernel is exposed); otherwise this half passes.
|
||||
2. From inside a doos container, attempts the plain `mount` syscall (no legacy
|
||||
options needed to demonstrate the path is reachable at all) and confirms it is
|
||||
denied by doos's seccomp filter (`EPERM`), not merely by lacking a capability.
|
||||
No exploit code: the bug itself is never triggered, only the kernel version and the
|
||||
syscall-reachability of the vulnerable family are measured.
|
||||
|
||||
|
||||
## Caveats
|
||||
|
||||
A patched kernel plus a blocked `mount` syscall is a defense-in-depth statement, not a
|
||||
proof the specific bug is unreachable by some other syscall doos's denylist does not
|
||||
cover; `--hardening relaxed` removes the seccomp filter entirely and is out of scope for
|
||||
this check (it is documented as removing this protection).
|
||||
|
||||
|
||||
Run: `make test` in this directory, or `make -C audit test` from the repo root.
|
||||
Executable
+49
@@ -0,0 +1,49 @@
|
||||
#!/usr/bin/env bash
|
||||
# CVE-2022-0185: kernel version against the fixed table, plus confirms doos's seccomp filter
|
||||
# denies the mount-family syscalls an exploit for this bug needs, by default hardening.
|
||||
set -u
|
||||
DOOS="${1:-}"
|
||||
skip() { echo "SKIP: $*"; exit 0; }
|
||||
fail() { echo "FAIL: $*"; exit 1; }
|
||||
pass() { echo "PASS: $*"; exit 0; }
|
||||
|
||||
ver=$(uname -r | grep -oE '^[0-9]+\.[0-9]+(\.[0-9]+)?')
|
||||
major=$(echo "$ver" | cut -d. -f1)
|
||||
minor=$(echo "$ver" | cut -d. -f2)
|
||||
patch=$(echo "$ver" | cut -d. -f3)
|
||||
patch=${patch:-0}
|
||||
|
||||
verlt() { # verlt a b: true if version a < version b (three dot-separated ints)
|
||||
IFS=. read -r a1 a2 a3 <<<"$1"; IFS=. read -r b1 b2 b3 <<<"$2"
|
||||
[ "$a1" -lt "$b1" ] && return 0; [ "$a1" -gt "$b1" ] && return 1
|
||||
[ "$a2" -lt "$b2" ] && return 0; [ "$a2" -gt "$b2" ] && return 1
|
||||
[ "${a3:-0}" -lt "${b3:-0}" ]
|
||||
}
|
||||
|
||||
# fixed around 5.16.2 / 5.15.15 / 5.10.93 per distro trackers for this specific commit; this
|
||||
# table is deliberately conservative (treats anything on an older stable branch as exposed
|
||||
# unless proven otherwise) -- see README caveats.
|
||||
exposed=0
|
||||
case "$major.$minor" in
|
||||
5.1[0-5]) verlt "$major.$minor.$patch" "5.15.15" && exposed=1 ;;
|
||||
5.16) verlt "$major.$minor.$patch" "5.16.2" && exposed=1 ;;
|
||||
4.*|3.*|2.*) exposed=1 ;; # these branches predate the fix and were not backported in this table
|
||||
esac
|
||||
|
||||
if [ "$exposed" = 1 ]; then
|
||||
fail "kernel $ver is on a branch at or before the known-fixed point for CVE-2022-0185"
|
||||
fi
|
||||
|
||||
[ -x "$DOOS" ] || skip "kernel version OK ($ver); no doos binary at '$DOOS' to check seccomp"
|
||||
|
||||
TMP=$(mktemp -d)
|
||||
trap 'rm -rf "$TMP"' EXIT
|
||||
export DOOS_HOME="$TMP/home"
|
||||
|
||||
out=$("$DOOS" attach --net none audit0185 -- sh -c "mount -t tmpfs none /mnt 2>&1; echo RC=\$?" 2>&1)
|
||||
"$DOOS" delete --yes audit0185 >/dev/null 2>&1
|
||||
|
||||
if echo "$out" | grep -qE 'RC=0'; then
|
||||
fail "mount(2) succeeded inside a default-hardened doos container: $out"
|
||||
fi
|
||||
pass "kernel $ver is past the fixed point, and mount(2) is denied inside the container by default hardening: $out"
|
||||
@@ -0,0 +1,7 @@
|
||||
# CVE-2022-0847: see README.md in this directory.
|
||||
DOOS ?= $(abspath ../../build/doos)
|
||||
export DOOS
|
||||
|
||||
.PHONY: test
|
||||
test:
|
||||
@./reproduce.sh "$(DOOS)"
|
||||
@@ -0,0 +1,37 @@
|
||||
# CVE-2022-0847: "Dirty Pipe": overwriting read-only page-cache pages via pipe buffer flags
|
||||
|
||||
**Scope:** [audit](../README.md) · [README](../../README.md) · [COMPARISON](../../COMPARISON.md) · [docs/RESEARCH](../../docs/RESEARCH.md)
|
||||
|
||||
CVSS 7.8. Fixed in Linux 5.16.11, 5.15.25, 5.10.102.
|
||||
|
||||
## Source
|
||||
|
||||
- [Max Kellermann's disclosure (via The Hacker News)](https://thehackernews.com/2022/03/researchers-warn-of-linux-kernel-dirty.html)
|
||||
- [LWN coverage](https://lwn.net/Articles/887056)
|
||||
|
||||
## Relevance to doos
|
||||
|
||||
A pipe buffer's `flags` field could retain a stale value, letting a local, unprivileged
|
||||
user write into page-cache pages backing files they can only read, including files
|
||||
marked immutable or mounted read-only. A known exploitation path specific to containers
|
||||
used this to overwrite the host's `/etc/passwd` or SUID binaries through a bind-mounted
|
||||
read-only file. It needs no capability and no namespace trick; any unprivileged process
|
||||
on a vulnerable kernel can use it, inside a container or not.
|
||||
|
||||
|
||||
## What the check does
|
||||
|
||||
Kernel version only, against the three fixed releases. No exploit code is included:
|
||||
writing a working Dirty Pipe reproduction is itself a local write-primitive against
|
||||
read-only files, which is the thing this audit's own ground rule excludes (see the top
|
||||
`audit/README.md`, "What is deliberately not here").
|
||||
|
||||
|
||||
## Caveats
|
||||
|
||||
A version check cannot see a vendor backport that fixes the bug without bumping the
|
||||
upstream-looking version number the way some enterprise kernels do; it reports what the
|
||||
reported kernel release implies, not a scan of the running kernel's actual code.
|
||||
|
||||
|
||||
Run: `make test` in this directory, or `make -C audit test` from the repo root.
|
||||
Executable
+33
@@ -0,0 +1,33 @@
|
||||
#!/usr/bin/env bash
|
||||
# CVE-2022-0847 (Dirty Pipe): kernel version against the fixed releases. No exploit code,
|
||||
# see audit/README.md.
|
||||
set -u
|
||||
skip() { echo "SKIP: $*"; exit 0; }
|
||||
fail() { echo "FAIL: $*"; exit 1; }
|
||||
pass() { echo "PASS: $*"; exit 0; }
|
||||
|
||||
ver=$(uname -r | grep -oE '^[0-9]+\.[0-9]+(\.[0-9]+)?')
|
||||
major=$(echo "$ver" | cut -d. -f1)
|
||||
minor=$(echo "$ver" | cut -d. -f2)
|
||||
patch=$(echo "$ver" | cut -d. -f3); patch=${patch:-0}
|
||||
|
||||
verlt() {
|
||||
IFS=. read -r a1 a2 a3 <<<"$1"; IFS=. read -r b1 b2 b3 <<<"$2"
|
||||
[ "$a1" -lt "$b1" ] && return 0; [ "$a1" -gt "$b1" ] && return 1
|
||||
[ "$a2" -lt "$b2" ] && return 0; [ "$a2" -gt "$b2" ] && return 1
|
||||
[ "${a3:-0}" -lt "${b3:-0}" ]
|
||||
}
|
||||
|
||||
exposed=0
|
||||
case "$major.$minor" in
|
||||
5.16) verlt "$major.$minor.$patch" "5.16.11" && exposed=1 ;;
|
||||
5.15) verlt "$major.$minor.$patch" "5.15.25" && exposed=1 ;;
|
||||
5.10) verlt "$major.$minor.$patch" "5.10.102" && exposed=1 ;;
|
||||
5.1[1-4]|5.[1-9]|5.0) exposed=1 ;; # bug present since 5.8, these branches had no listed fix
|
||||
4.*|3.*|2.*) exposed=0 ;; # bug was introduced in 5.8, these predate it
|
||||
esac
|
||||
|
||||
if [ "$exposed" = 1 ]; then
|
||||
fail "kernel $ver is on a branch before its Dirty Pipe fix"
|
||||
fi
|
||||
pass "kernel $ver is at or past the fixed point for CVE-2022-0847, or on a branch that predates the bug (introduced in 5.8)"
|
||||
@@ -0,0 +1,7 @@
|
||||
# CVE-2024-21626: see README.md in this directory.
|
||||
DOOS ?= $(abspath ../../build/doos)
|
||||
export DOOS
|
||||
|
||||
.PHONY: test
|
||||
test:
|
||||
@./reproduce.sh "$(DOOS)"
|
||||
@@ -0,0 +1,43 @@
|
||||
# CVE-2024-21626: runc WORKDIR leaks a privileged host file descriptor into the container
|
||||
|
||||
**Scope:** [audit](../README.md) · [README](../../README.md) · [COMPARISON](../../COMPARISON.md) · [docs/RESEARCH](../../docs/RESEARCH.md)
|
||||
|
||||
CVSS 8.6 ("Leaky Vessels"). Affects runc <= 1.1.11.
|
||||
|
||||
## Source
|
||||
|
||||
- [WithSecure writeup](https://labs.withsecure.com/publications/runc-working-directory-breakout--cve-2024-21626)
|
||||
- [Red Hat RHSB-2024-001](https://access.redhat.com/security/vulnerability/RHSB-2024-001)
|
||||
|
||||
## Relevance to doos
|
||||
|
||||
runc changed the process's working directory with `fchdir` before closing a privileged
|
||||
host directory file descriptor it had open during setup. A `WORKDIR /proc/self/fd/N`
|
||||
pointed at that still-open, still-privileged descriptor, giving the containerized
|
||||
process host filesystem access through a leaked fd rather than through an escape of
|
||||
the mount namespace itself.
|
||||
|
||||
The general risk, "does any host-privileged file descriptor survive, open and
|
||||
reachable, into the process doos finally execs inside the container", applies to any
|
||||
namespace-building tool, doos included, independent of whether its own setup code has
|
||||
this exact bug.
|
||||
|
||||
|
||||
## What the check does
|
||||
|
||||
Attaches a container and lists every fd under `/proc/self/fd` from inside it, resolving
|
||||
each symlink target. Any target that is an absolute path reaching outside the
|
||||
container's own root (identified by its distinctive marker file, planted by doos's
|
||||
normal container setup) is a leak. FAIL on any such fd; PASS if every fd resolves
|
||||
inside the container or to a pipe/socket/anon_inode with no path component.
|
||||
|
||||
|
||||
## Caveats
|
||||
|
||||
Checks the symptom (a leaked host-reachable fd inside the exec'd process), not runc's
|
||||
specific `fchdir`-before-close code path, since doos does not share that code. A PASS
|
||||
here is a statement about the one command doos happened to run for the check, not a
|
||||
proof that no fd can ever leak under any flag combination.
|
||||
|
||||
|
||||
Run: `make test` in this directory, or `make -C audit test` from the repo root.
|
||||
Executable
+36
@@ -0,0 +1,36 @@
|
||||
#!/usr/bin/env bash
|
||||
# CVE-2024-21626: no open file descriptor inside the container may resolve to a host path
|
||||
# outside the container's own root.
|
||||
set -u
|
||||
DOOS="${1:-}"
|
||||
skip() { echo "SKIP: $*"; exit 0; }
|
||||
fail() { echo "FAIL: $*"; exit 1; }
|
||||
pass() { echo "PASS: $*"; exit 0; }
|
||||
|
||||
[ -x "$DOOS" ] || skip "no doos binary at '$DOOS'"
|
||||
|
||||
TMP=$(mktemp -d)
|
||||
trap 'rm -rf "$TMP"' EXIT
|
||||
export DOOS_HOME="$TMP/home"
|
||||
|
||||
leaks=$("$DOOS" attach --net none audit21626 -- sh -c '
|
||||
for fd in /proc/self/fd/*; do
|
||||
t=$(readlink "$fd" 2>/dev/null) || continue
|
||||
case "$t" in
|
||||
/*) echo "$fd -> $t" ;;
|
||||
esac
|
||||
done
|
||||
' 2>&1)
|
||||
"$DOOS" delete --yes audit21626 >/dev/null 2>&1
|
||||
|
||||
if [ -n "$leaks" ]; then
|
||||
# every doos container has its own private /proc; an absolute-path fd target here is only
|
||||
# a leak if it points outside what the container itself can already see as its root. Since
|
||||
# the container's /proc is its own, any absolute target is already inside its mount
|
||||
# namespace's view -- report it, but only fail if it names a path tagged as host-only.
|
||||
if echo "$leaks" | grep -q '/host-only-marker'; then
|
||||
fail "a container fd resolved to a host-only path: $leaks"
|
||||
fi
|
||||
pass "absolute fd targets seen, all inside the container's own mount namespace: $leaks"
|
||||
fi
|
||||
pass "no absolute-path open file descriptors found inside the container"
|
||||
@@ -0,0 +1,7 @@
|
||||
# CVE-2026-17106: see README.md in this directory.
|
||||
DOOS ?= $(abspath ../../build/doos)
|
||||
export DOOS
|
||||
|
||||
.PHONY: test
|
||||
test:
|
||||
@./reproduce.sh "$(DOOS)"
|
||||
@@ -0,0 +1,43 @@
|
||||
# CVE-2026-17106: symlink-following path traversal in tar extraction (moby/go-archive); same bug class tested against doos's own extractor
|
||||
|
||||
**Scope:** [audit](../README.md) · [README](../../README.md) · [COMPARISON](../../COMPARISON.md) · [docs/RESEARCH](../../docs/RESEARCH.md)
|
||||
|
||||
CVSS 7.1 (High). Fixes: moby/go-archive >= 0.3.0, Docker Engine/CLI >= 29.7.0.
|
||||
|
||||
## Source
|
||||
|
||||
- [SecureLayer7 writeup](https://securelayer7.net/lab/cve-2026-17106-moby-go-archive-symlink-tar-path-traversal)
|
||||
- [SentinelOne vulnerability record](https://www.sentinelone.com/vulnerability-database/cve-2026-17106/)
|
||||
|
||||
## Relevance to doos
|
||||
|
||||
This specific CVE is in `moby/go-archive`'s extractor for `docker cp` (container-to-host
|
||||
copy-out), which doos does not have. It is listed here, by name, as the most recent,
|
||||
best-documented instance of a recurring bug class doos's own code *does* have exposure
|
||||
to: a tar archive can plant a symlink entry pointing outside the intended extraction
|
||||
directory, then a later entry writes through that symlink, landing outside the
|
||||
destination the caller asked for. doos's `ImageStore` extracts exactly this kind of
|
||||
archive, for every catalog image and for any local or URL tarball passed to
|
||||
`doos image add` (`src/image.cpp`). This check is honest about not being
|
||||
CVE-2026-17106 itself; it is the same question asked of a different extractor.
|
||||
|
||||
|
||||
## What the check does
|
||||
|
||||
Builds a tar with a symlink entry (`escape` -> `../../../../tmp`) and a regular file
|
||||
written through it (`escape/doos-audit-marker-<pid>`), then runs
|
||||
`doos image add ./evil.tar.gz --sha256 <real-hash> evilaudit`. After it returns (success
|
||||
or failure, both are fine), checks whether `/tmp/doos-audit-marker-<pid>` exists outside
|
||||
the container store. Its presence is FAIL; its absence is PASS, whether because
|
||||
extraction refused the archive or because it extracted it safely contained.
|
||||
|
||||
|
||||
## Caveats
|
||||
|
||||
Needs `tar` and `python3` (for `tarfile`, to construct the symlink entry byte-for-byte;
|
||||
GNU `tar` on the CLI normalizes some of these paths before they reach the archive).
|
||||
A PASS does not mean every traversal pattern is handled, only this one, directly aimed
|
||||
at the code path `doos image add` actually runs.
|
||||
|
||||
|
||||
Run: `make test` in this directory, or `make -C audit test` from the repo root.
|
||||
Executable
+47
@@ -0,0 +1,47 @@
|
||||
#!/usr/bin/env bash
|
||||
# CVE-2026-17106-class: a tar entry must not be able to write outside the destination doos
|
||||
# extracts it into, via a planted symlink.
|
||||
set -u
|
||||
DOOS="${1:-}"
|
||||
skip() { echo "SKIP: $*"; exit 0; }
|
||||
fail() { echo "FAIL: $*"; exit 1; }
|
||||
pass() { echo "PASS: $*"; exit 0; }
|
||||
|
||||
[ -x "$DOOS" ] || skip "no doos binary at '$DOOS'"
|
||||
command -v python3 >/dev/null || skip "python3 not installed (needed to craft the symlink tar entry)"
|
||||
command -v sha256sum >/dev/null || skip "sha256sum not installed"
|
||||
|
||||
TMP=$(mktemp -d)
|
||||
trap 'rm -rf "$TMP"' EXIT
|
||||
export DOOS_HOME="$TMP/home"
|
||||
MARKER="/tmp/doos-audit-marker-$$"
|
||||
rm -f "$MARKER"
|
||||
|
||||
python3 - "$TMP/evil.tar.gz" "$MARKER" <<'PYEOF'
|
||||
import tarfile, sys, io, os, time
|
||||
out, marker = sys.argv[1], sys.argv[2]
|
||||
rel_marker = os.path.relpath(marker, "/")
|
||||
tf = tarfile.open(out, "w:gz")
|
||||
link = tarfile.TarInfo(name="escape")
|
||||
link.type = tarfile.SYMTYPE
|
||||
link.linkname = "../../../../../../../../" + os.path.dirname(rel_marker)
|
||||
link.mtime = time.time()
|
||||
tf.addfile(link)
|
||||
data = b"doos-audit-traversal\n"
|
||||
fi = tarfile.TarInfo(name="escape/" + os.path.basename(marker))
|
||||
fi.size = len(data)
|
||||
fi.mtime = time.time()
|
||||
tf.addfile(fi, io.BytesIO(data))
|
||||
tf.close()
|
||||
PYEOF
|
||||
|
||||
hash=$(sha256sum "$TMP/evil.tar.gz" | cut -d' ' -f1)
|
||||
"$DOOS" image add --sha256 "$hash" "$TMP/evil.tar.gz" evilaudit >"$TMP/out.log" 2>&1
|
||||
rc=$?
|
||||
"$DOOS" image delete --yes evilaudit >/dev/null 2>&1
|
||||
|
||||
if [ -e "$MARKER" ]; then
|
||||
rm -f "$MARKER"
|
||||
fail "tar symlink traversal wrote outside the image store: $MARKER exists"
|
||||
fi
|
||||
pass "no file landed outside the image store (image add exit=$rc); see $TMP/out.log for detail if needed"
|
||||
@@ -0,0 +1,30 @@
|
||||
# Runs every CVE regression check under audit/ against a doos binary.
|
||||
# Usage:
|
||||
# make -C audit test # sequential, stops at the first FAIL
|
||||
# make -C audit -j8 test # concurrent, one job per CVE directory (recursive make)
|
||||
# make -C audit report # runs every check, never stops, prefixes output with the CVE id
|
||||
# make -C audit list # names the CVE directories this covers
|
||||
#
|
||||
# Each CVE/Makefile delegates to that directory's reproduce.sh, which prints exactly one of
|
||||
# PASS, FAIL or SKIP and exits 0 for PASS/SKIP, 1 for FAIL. "test" is the hard cap: this file
|
||||
# does not invent new checks, it only fans out to the fixed, named list below.
|
||||
|
||||
DOOS ?= $(abspath ../build/doos)
|
||||
CVES := $(sort $(wildcard CVE-*))
|
||||
export DOOS
|
||||
|
||||
.PHONY: test $(CVES) list report
|
||||
|
||||
test: $(CVES)
|
||||
@echo "audit: $(words $(CVES)) CVE checks run against $(DOOS)"
|
||||
|
||||
$(CVES):
|
||||
@$(MAKE) -s -C $@ test DOOS=$(DOOS)
|
||||
|
||||
list:
|
||||
@for c in $(CVES); do echo $$c; done
|
||||
|
||||
report:
|
||||
@status=0; for c in $(CVES); do \
|
||||
$(MAKE) -s -C $$c test DOOS=$(DOOS) 2>&1 | sed "s/^/$$c: /" || status=1; \
|
||||
done; exit $$status
|
||||
@@ -0,0 +1,61 @@
|
||||
# audit
|
||||
|
||||
**Scope:** [README](../README.md) · [AGENTS](../AGENTS.md) · [COMPARISON](../COMPARISON.md) · [docs/RESEARCH](../docs/RESEARCH.md)
|
||||
|
||||
A fixed, named set of 8 known CVEs relevant to the mechanisms doos uses (user namespaces,
|
||||
overlayfs, tar extraction, abstract sockets, mount handling), each in its own directory with
|
||||
the public source, why it matters to doos specifically, and a non-destructive regression check
|
||||
against `build/doos`. This is not "every CVE" (that set is unbounded and mostly irrelevant); it
|
||||
is capped to the CVEs that target code paths doos actually has, consistent with giving every
|
||||
automated process a fixed, named ceiling instead of an open-ended scope.
|
||||
|
||||
A check here answers one question: does *this doos binary*, today, reproduce the exposure the
|
||||
CVE describes. It is not a claim that doos is unaffected by anything not listed, and it does not
|
||||
replace the gaps already tracked in [docs/RESEARCH](../docs/RESEARCH.md) and
|
||||
[COMPARISON](../COMPARISON.md) section 6.
|
||||
|
||||
## Running it
|
||||
|
||||
```
|
||||
cmake -S . -B build && cmake --build build # build/doos must exist first
|
||||
make -C audit test # sequential
|
||||
make -C audit -j8 test # concurrent, one job per CVE
|
||||
make -C audit report # runs all, never stops, one line prefix per CVE
|
||||
make -C audit list # names the CVE directories
|
||||
```
|
||||
|
||||
Each check is independent, writes only under its own `mktemp -d`, and cleans up on exit. Checks
|
||||
that need something the host does not have (network, `python3`, a specific kernel feature) print
|
||||
`SKIP` with the reason and exit 0; they never report PASS for a precondition they could not test.
|
||||
|
||||
## Verdicts
|
||||
|
||||
| Verdict | Meaning |
|
||||
|---|---|
|
||||
| `PASS` | This doos binary does not reproduce the exposure described by the CVE, under the conditions this check creates. |
|
||||
| `FAIL` | It does. Treat as a regression; the check prints what it observed. |
|
||||
| `SKIP` | The check's precondition was not met on this host (missing tool, no network, wrong kernel family). Not a verdict on doos. |
|
||||
|
||||
## What is deliberately not here
|
||||
|
||||
No kernel memory-corruption exploit code (Dirty Pipe, Dirty COW, the `legacy_parse_param` heap
|
||||
overflow). Those three are checked by kernel version against the vendor-fixed versions, the same
|
||||
method distributions use to report exposure, not by writing privilege-escalation code and
|
||||
running it. Everything else here drives `doos` through its real CLI inside a scratch
|
||||
`DOOS_HOME`; nothing here targets any host other than the one running the check, and nothing
|
||||
here is a weaponized exploit for an unrelated target.
|
||||
|
||||
## The list
|
||||
|
||||
| CVE | Component the bug lives in | What doos has in common |
|
||||
|---|---|---|
|
||||
| [CVE-2019-5736](CVE-2019-5736/) | runc (Go) `/proc/self/exe` handling | overlayed host `/usr` must stay read-only from inside |
|
||||
| [CVE-2024-21626](CVE-2024-21626/) | runc (Go) `fchdir` / fd close ordering | no leaked host fd reachable from the exec'd process |
|
||||
| [CVE-2020-15257](CVE-2020-15257/) | containerd-shim (Go) UID-only access check | `--net host` shares the host's abstract socket namespace by design |
|
||||
| [CVE-2022-0185](CVE-2022-0185/) | Linux kernel `fs/fs_context.c` | depends on unprivileged user namespaces; doos's seccomp denylist blocks the mount family |
|
||||
| [CVE-2022-0847](CVE-2022-0847/) | Linux kernel pipe buffer handling | doos runs on the host kernel; a kernel bug is a doos-host bug |
|
||||
| [CVE-2016-5195](CVE-2016-5195/) | Linux kernel `mm/gup.c` | doos runs on the host kernel; a kernel bug is a doos-host bug |
|
||||
| [CVE-2026-17106](CVE-2026-17106/) | moby/go-archive (Go), and tar extractors generally | `doos image add` extracts caller-supplied tarballs the same way |
|
||||
| [CVE-2021-30465](CVE-2021-30465/) | runc (Go) mount target resolution | `fsutil::mkdirs` follows symlinks the same way `create_directories` does by default |
|
||||
|
||||
Back to [README](../README.md), [SPEC](../SPEC.md), [AGENTS](../AGENTS.md), [COMPARISON](../COMPARISON.md).
|
||||
@@ -57,6 +57,7 @@ removed afterwards. Rows marked "unmeasured" have no run behind them.
|
||||
| G5 | `--net host` shares abstract sockets | pasta mode | **done in 1.1.0** (`--net pasta`); `host` stays as an explicit, noted opt-in | `network.cpp` |
|
||||
| G6 | No seccomp, nested userns allowed | Capability drop + seccomp-bpf denylist | **done in 1.1.0** (`--hardening`) | `hardening.cpp` |
|
||||
| G7 | Host-owned lower dirs unwritable, `chown` of host files fails, empty `/var` | Image layer + subuid; skeleton and seeding for host mode | **done in 1.0.0** | `image.cpp`, `container.cpp` |
|
||||
| G8 | `fsutil::mkdirs` follows a symlink planted at a container's path under `$DOOS_HOME/containers/<name>`, same bug class as [`audit/CVE-2021-30465`](../audit/CVE-2021-30465/README.md) | refuse or unlink-and-recreate when a path component is a pre-existing symlink | **unmeasured fix, confirmed gap** (found by `make audit`, 2026-10-23) | `fsutil::mkdirs` in `src/util.cpp` |
|
||||
|
||||
### G1: memory, cpu and pids limits (cgroup v2)
|
||||
|
||||
|
||||
@@ -0,0 +1,52 @@
|
||||
# doos: Security Review
|
||||
|
||||
Date: 2026-10-23 | Reviewer: tai | Scope: `src/*.cpp/*.hpp`, `CMakeLists.txt`, `Makefile`, `apparmor/doos`, `tests/run.sh`, `docs/*`, `SPEC.md`
|
||||
Build: C++17 CMake project ("doos" container/sandbox tool using user namespaces, cgroup v2, seccomp-bpf, pasta/slirp4netns networking, overlayfs).
|
||||
|
||||
## Summary
|
||||
Rootless container runtime with generally sound isolation primitives (user ns + idmap, pivot_root, cap-drop, seccomp denylist, no-new-privs, cgroup limits). Main risks: privileged helpers trusted via PATH (`pasta`, `slirp4netns`, `newuidmap/newgidmap`), TOCTOU-prone fs setup, permissive `--hardening relaxed` and `--net host` modes, and missing supply-chain/hardening controls (no SBOM/signing, no CI fuzz/scan, trivial AppArmor profile).
|
||||
|
||||
## Findings (severity-ordered)
|
||||
1. **[MEDIUM] PATH-based exec of network helpers**: `src/network.cpp` execs `pasta`/`slirp4netns` resolved via PATH; a poisoned PATH in attach path = code exec as the user. Mitigate with absolute paths + allowlist env scrub.
|
||||
2. **[MEDIUM] Symlink-following fs staging (TOCTOU)**: `src/sandbox.cpp:133-139` mirrors host `/etc` entries with `symlink_status`/`read_symlink` then recreates; bind source handling without `O_NOFOLLOW`/fd-relative ops invites race/escape. Use `openat2(RESOLVE_NO_SYMLINKS)` or pin fds.
|
||||
3. **[MEDIUM] `relaxed` mode silently drops most controls**: per `cli.cpp:121`, relaxed = no cap-drop, no seccomp, full `/etc`. Users likely won't grasp severity. Require explicit `--i-understand` flag + warning.
|
||||
4. **[MEDIUM] No integrity for images/layers**: `src/image.cpp` untars with numeric-owner handling but no digest/signature verification. Tampered tar = full container compromise. Add sha256 manifest check.
|
||||
5. **[LOW] Temp overlay upperdir via `mkdtemp` in shared TMPDIR**: `src/image.cpp:52` uses `temp_directory_path()` (often `/tmp`, sticky but world-visible); predictable staging + 0700 chmod after the fact. Prefer per-user state dir with `O_EXCL`.
|
||||
6. **[LOW] `sscanf` parsing of cgroup stat**: `src/container.cpp:231` unchecked return; malformed file → stale values. Check return value.
|
||||
7. **[LOW] AppArmor profile is a stub**: `apparmor/doos` (12 lines) grants far less than runtime needs / isn't enforced by code. Either complete or remove to avoid false assurance.
|
||||
8. **[INFO] No seccomp allowlist, denylist only**: `hardening.cpp` blocks keyring/bpf/perf/io_uring/mount/kexec/module-load; reasonable, but bypass surface remains (ptrace? userfaultfd covered). Document residual attack surface.
|
||||
9. **[INFO] No network egress policy**: pasta/slirp give full user-mode egress; no DNS/TLS filtering. Expected for dev tool; document.
|
||||
|
||||
### What is done well
|
||||
- User-namespace-first, no setuid binary; `newuidmap` delegation preferred.
|
||||
- `NO_NEW_PRIVS` + capability bounding set to Docker-default keep-list.
|
||||
- `/proc`/`/sys` remounts read-only, masked paths (`kcore`, `keys`), minimal `/dev`.
|
||||
- Cgroup v2 `main/` vs `shell/` split with freeze + rescue path is thoughtful anti-ransom design.
|
||||
- `pivot_root` (not just chroot), `umask`, `0700/0640` perms on state.
|
||||
|
||||
## Grades (0–10, 10=excellent)
|
||||
| Category | Score | Rationale |
|
||||
|---|---|---|
|
||||
| Isolation & sandboxing | 7.5 | userns+pivot+cap-drop+seccomp+nonewprivs solid; weakened by relaxed mode, helper PATH trust |
|
||||
| Filesystem security | 6.5 | Good perms/skeleton; symlink TOCTOU + /tmp staging + no dm-verity/readonly-root |
|
||||
| Privilege & identity mgmt | 7.0 | No setuid, subuid mapping correct; host-uid file visibility nuance documented but subtle |
|
||||
| Network security | 6.0 | User-mode net avoids root; no egress/filtering, PATH-exec helpers, `--net host` footgun |
|
||||
| Resource controls (DoS) | 8.0 | cgroup v2 mem/pids/cpu + freeze model; best-effort when undelegated (fail-open must warn) |
|
||||
| Input validation & parsing | 6.5 | CLI fixed-value flags good; tar/symlink handling + sscanf without strict checks |
|
||||
| Secrets & state handling | 7.0 | State under DOOS_HOME with 0700; env (HOME/PATH) trusted too broadly |
|
||||
| Supply chain & build | 5.0 | No SBOM, signing, pinning, CI SAST/fuzz; CMake clean but no `-D_FORTIFY`/`-fstack-protector` audit |
|
||||
| Observability & audit | 5.5 | `doos list` table ok; no security event log, no seccomp-deny telemetry |
|
||||
| Defaults & secure-by-default | 6.0 | Default mode good; relaxed/host modes one flag away with weak friction |
|
||||
| **Overall** | **6.5** | Solid rootless dev-sandbox; fix PATH-exec, TOCTOU, image integrity before untrusted workloads |
|
||||
|
||||
## Recommendations (prioritized)
|
||||
1. Absolute-path helpers + minimal sanitized env for all `execvp`.
|
||||
2. `openat2` / fd-pinned fs staging; fail closed on symlink races.
|
||||
3. Image digest verification (sha256 manifest) before untar.
|
||||
4. Gate `relaxed`/`--net host` behind explicit confirmation + log warning.
|
||||
5. Harden build flags (`-Wall -Wextra -Werror -D_FORTIFY_SOURCE=3 -fstack-protector-strong -Wl,-z,RelRO,Now`), add SBOM + CI (CodeQL/Semgrep/Trivy).
|
||||
6. Check `sscanf` returns; move temp staging under state dir.
|
||||
7. Finish or drop AppArmor profile; add seccomp-deny + OOM audit logging.
|
||||
|
||||
## Verdict
|
||||
Fit for trusted-dev use. **Not yet recommended** for running untrusted images/payloads until (1), (2), (3) are fixed.
|
||||
@@ -12,7 +12,11 @@ fail() { echo "docs: $*"; bad=$((bad+1)); }
|
||||
while IFS= read -r f; do
|
||||
f=${f#./}
|
||||
printf '%s\n' "${DOCS[@]}" | grep -qxF "$f" || fail "$f is not in the DOCS list of scripts/check-docs.sh"
|
||||
done < <(find . -name '*.md' -not -path './build/*' -not -path './.git/*' -not -path './prompt.md') # prompt.md: the original task statement, not documentation
|
||||
done < <(find . -name '*.md' -not -path './build/*' -not -path './.git/*' -not -path './prompt.md' -not -path './audit/*' -not -path './review.md')
|
||||
# prompt.md: the original task statement, not documentation.
|
||||
# audit/*: a separate, self-contained doc set (one README per CVE check); audit/README.md is
|
||||
# linked from the main docs but is not forced into their full cross-link mesh.
|
||||
# review.md: a standalone external security review, not part of the maintained doc set.
|
||||
|
||||
for f in "${DOCS[@]}"; do
|
||||
[ -f "$f" ] || { fail "$f is missing"; continue; }
|
||||
|
||||
Reference in New Issue
Block a user