MongoDB Community · Isolated lab · v6 evaluation
Stop ransomware by
how much it can hurt,
not just that it exists.
AurevoDB watches MongoDB for destructive commands, scores the damage in real time, and contains the attack in graded steps — throttle, suspend, isolate — then restores what was lost. No binary alarm. A measured response.
95% CI [44.99, 52.49], p=0.0001
How it works
One score. Four responses. No guesswork.
Every change-stream event feeds a damage-weighted score. The score — not a single rule — decides what happens next.
Watch
Score is tracked. Nothing is blocked. Normal traffic — bulk inserts, index builds, backups — stays under 30.
Throttle + snapshot
Attacker sessions are throttled (killSessions) and a snapshot is taken on suspicion, so there is something to restore from.
Suspend + quarantine
Writes are suspended and the attacker's role is revoked on the authenticated server — later batches fail with code 13.
Isolate + restore
The database is write-locked (fsyncLock) and the snapshot is restored — staging copy first, rename-swap only on success, verified against recorded counts.
drop() is not killable — killOp acknowledged, the drop still completed. killSessions interrupts one operation, not the session. So AurevoDB does not promise to kill a drop. It promises to detect it fast, limit what follows, and restore what was lost. That is the whole design.
Live Lab
Run the attack yourself.
Pick an attack shape and a mode. The lab simulates the seeded database the v6 evaluation used — shop 10,000 docs with payments topped up to 20,000 (26,500 total) — scores every event with the real formula, and fires the real graded actions at 30 / 60 / 85.
Interactive simulation of the AurevoDB algorithm in your browser. Outcomes are calibrated to the v6 lab evaluation medians — this is not a live MongoDB server, and it is not the Python engine.
Graded: throttle at 30, suspend at 60, isolate at 85, then restore.
Run result
Results — v6 evaluation
The numbers, all of them.
426 attack runs + 48 benign runs + 18 restore runs, fresh seeds 11 / 42 / 777, on the authenticated mongod by default. Medians per shape, graded containment vs detect-only (which saved 0% everywhere).
| Attack shape | Saved by prevention — median | Mean | Zero-save runs |
|---|---|---|---|
| Fast drop | 71.92% | 71.47% | 0 / 45 |
| Fast delete | 54.08% | 48.89% | 3 / 45 |
| Slow delete | 68.00% | 68.00% | 0 / 27 |
| Sustained | 60.00% | 58.52% | 0 / 27 |
| Interleaved | 32.60% | 30.72% | 1 / 27 |
| dropdb | 0.00% | 0.00% | 27 / 27 — by design |
95% CI [44.99, 52.49], p=0.0001, n=198
95% CI [16.52, 43.24], p=0.0008, n=18 — first fully-powered win
Progression across versions
Pooled graded-vs-detect-only points per version. v5 pooled only the two fast shapes (+42.74) — it is not comparable to the all-six-shape numbers. v6 pools all six shapes again.
Restore, verified end to end
median restore 2,851.4 ms
median restore 2,669.2 ms
Honest limits
What it does not do.
A product page that hides its limits is a brochure. These are measured, and they stay in the report.
- dropdb prevention is 0% by design. A
dropDatabasecannot be killed. The answer is verified fast restore (median 2,669.2 ms) plus least-privilege roles, not prevention. - Steady overhead is +242.8% mongod CPU. The ≤+43% target was proven unreachable: priming off still costs +49.1%, and only the 0.02 s priming rate keeps fast-delete alive.
- Undeclared maintenance looks like an attack. A bulk-delete purge trips the alarm 3/3 times, and an orders drop-and-recreate 3/3 times. Declared maintenance is exempt and scores 0/18 false positives — undeclared work is not.
- Finding 10 is bounded, not solved. Fast-delete delivery stalls move between seeds (v4: seed 42, v5: seed 777). The stall decomposes into majority-commit visibility and pure change-stream lag inside the server; the 8-minute clustering never reproduced.
- Damage-prevented % is an upper bound. The emulator is not validated against real ransomware traces, and the lab baselines flatter the system. The report states this in full.
- Interleaved attacks are the weakest shape at 32.60% median — mixed, pausing attacks are the hardest to contain.
Docs & report
Everything is open to check.
The full evaluation report is 37,052 words with a 491-word abstract — every headline number in it traces back to a results file. The lab code, evaluation scripts and this website are public.