Difficulty: 3.5/5 (held after calibration) · Docker Compose · 2 services
A timing side-channel with no code leak: the API never throws, never 500s, never tells you why it said no. You get HTTP codes and milliseconds — recover a secret from the difference.
AccessGate ("the gate") guards a facility with a four-segment numeric token. The gate checks segments sequentially and short-circuits — and ops refuses to add jitter because "the variance is acceptable". The internal beacon service holds the master secret. You are outside the fence with a stopwatch.
./RESET.sh # generates gate secret + seeds flags
# gate API: http://127.0.0.1:18920 (bind 0.0.0.0)Endpoints (self-documenting):
| route | what |
|---|---|
GET /status |
token-segments required, latency budget — free recon |
POST /gate |
probe a candidate (token=12-34-56-7890) |
POST /audit |
authenticated — needs the master secret |
GET /metrics |
response-time histogram (free) |
- Learn how the gate leaks time (segment-wise? whole-token? how many ms?).
- Recover the master secret well enough to pass
POST /audit. - Read
/opt/beacon/flag.txt(mounted inside the beacon service — reachable once you can drive the audit-side surface) — the flag file is what you hand in.
tools/dump_samples.py — high-rate latency sampler with per-code
separation; tools/sample_stats.py — median/MAD trimming and peak
analysis. (These are the same instruments used in the calibration run.)
./RESET.sh — regenerates the gate secret deterministically, re-seeds
flags, health-gates on /status.
The vulnerability is structural, not simulated: secrets.compare_digest
for the real comparison, time.perf_counter budgets, and an nginx
front that never buffers the timing signal away. The double-peak you will
find in the histogram is exactly where you expect it — the hard part is
recovering the tail (last segment) against noise, at a 70 %
false-positive budget.
MIT. Part of the <target>-lab pentest series — field-calibrated by an AI
pentest agent (see DIFFICULTY.md).