Part 6 of 13 · 1 min read
What the example found
Wrongly-allowed attacks per check design, and the resource-limit probes that actually ran.
How to read the checks table: each row is a check design against its own attack set; the second column is how often it wrongly let an attack through, the first how often it correctly allowed legitimate input.
| Check | Legitimate allowed | Attacks wrongly allowed |
|---|---|---|
Paths: startswith(root) |
100% | 80% (8 of 10) |
| Paths: normalise only | 100% | 30% (3 of 10) |
Paths: realpath + commonpath |
100% | 0% |
| URLs: "contains" | 100% | 80% |
| URLs: string prefix | 33% | 30% |
| URLs: parsed host allowlist | 100% | 0% |
Commands: startswith |
100% | 100% (9 of 9) |
| Commands: program and argument allowlist | 100% | 0% |
What to notice: the strong check in every category (resolve-then-compare for paths, parse-then-compare for URLs, allowlist-then-match for commands) reaches 0% wrongly allowed while losing nothing on legitimate input. There's no inherent trade-off between safety and usability here, only a trade-off between a cheap string check and a slightly more careful one.
How to read the probes table: each row is a resource-limited child process actually run on this machine.
| Probe | Observed result |
|---|---|
while True: pass |
killed at 1.02 s (CPU limit 1 s) |
2 GB bytearray |
MemoryError (limit 256 MB) |
| 50 MB file write | File too large (limit 1 MB) |
sleep(60) |
timed out at 5 s (wall clock) |
Read SECRET_TOKEN |
"not visible" scrubbed; "hunter2" inherited |
What to notice: every one of these five probes is a different kind of runaway behaviour (CPU, memory, disk, wall-clock, information disclosure). Each is stopped by a different, independent setting. There is no single "sandbox switch." Each row is its own control that has to be turned on deliberately.
{{svg:layers|Four layers: tool arguments, process limits, isolation boundary, network egress.}} {{svg:results|Wrongly allowed attack inputs per check, and the probes that ran.}}