loading...
IDS ROC curve: true-positive rate against false-positive rate (dot = your threshold)
Expected offline crack time (seconds, log scale) against password length; grey line = 1 year
running...
How to use
What it is. A small company network under steady hostile pressure: internet, edge firewall, DMZ web tier, application tier, data tier, office workstations and an IDS/SIEM sensor. Colours: green clean, red compromised, cyan isolated, amber alert pending.
Controls. Pick a preset, then move the sliders: prevention layers, sensor quality and threshold, analyst speed, threat rate and the password policy. Click a host to quarantine or restore it yourself. Hostile wave injects six extra attempts; New seed replays the same settings with different luck.
Try this. (1) Run Open flat network, then Defence in depth, and compare exposed data and the graph. (2) With defence in depth, drag the threshold from 3 down to 1 and watch false quarantines and availability fall (alert fatigue). (3) Set segmentation to 0 against 1 and see how far a single foothold spreads. (4) In the password panel, compare 6 digits against 14 mixed characters on a fast hash and on a memory-hard KDF.
Drag to orbit, wheel to zoom. Space pauses, R resets.
3D: a Blender 5.2 data hall (named racks, firewall wall, glass segmentation panes, SOC wall) baked to PBR textures and driven live by the same model. Drag to orbit, wheel to zoom, click a rack to isolate it.
The science
Defensive education only. Nothing here contains exploit code, payloads or step-by-step intrusion methods. An "attack" is an abstract hostile attempt that either fails or leaves one machine in a generic compromised state; the lab is about how controls change the odds. All probabilities are illustrative model coefficients, not measurements of any real product or incident.
Time and state. One model step is 5 simulated minutes. Every host is clean, compromised, or isolated. Hostile attempts arrive as a Poisson process with rate λ per hour, 55 % aimed at the public tier (load balancer, web servers) and 45 % at office workstations.
Prevention. Chance an attempt gains a foothold:
public tier: p = 0.55 (1 - 0.85 d)(1 - 0.8 u) + ponline workstation: p = 0.40 (1 - 0.8 a)(1 - 0.7 m)
with d = firewall default-deny coverage, u = patch coverage, a = awareness training, m = MFA coverage. Rebuilt (hardened) hosts keep only 30 % of that susceptibility.
Lateral movement and segmentation. Each compromised, connected host tries every link to a clean neighbour once per step. The link probability is a geometric (memoryless) process:
plink = β (1 - 0.75 u) · (1 - 0.92 s if the link crosses a zone boundary), β = 0.10 per step, s = segmentation
so the waiting time to spread over one link has mean 1/p. Identity and database hosts get a further (1 - 0.6 m) from MFA. Check 4 below runs the real stepping code on a two-host network and compares the mean waiting time with 1/p.
IDS as a Gaussian detector. Each connected host produces an anomaly score every step: N(0, 1) when clean and N(d′, 1) when compromised. An alert fires when the score exceeds the threshold τ:
FPR = Q(τ) = ½ erfc(τ/√2), TPR = Q(τ - d′), Q(x) = 1 - Φ(x)
Moving τ slides the operating point along the ROC curve. Because most hosts are clean, even a small FPR dominates the alert queue (the base-rate effect): with 16 hosts and τ = 1 you expect about 2.5 false alerts per step, which is exactly how "alert fatigue" and needless quarantines (an availability cost) arise. Checks 1 and 2 verify the simulated detector against the analytic Q function.
Response. An alert enters the analyst queue; after the response delay the host is isolated. A compromised host is rebuilt after 8 steps (40 min) and returns hardened; a falsely quarantined clean host is released after 3 steps. Clicking a host isolates it by hand.
Passwords and hashing. For a password drawn uniformly from an alphabet of S symbols and length L:
H = L log2 S bits, E[guesses] = 2H-1, E[time] = 2H-1 / r
where r is the guess rate. A fast unsalted hash allows enormous r on a GPU rig; salted slow hashes and memory-hard key-derivation functions lower r by orders of magnitude, which buys time without changing H. If the primary database is compromised, the model lets stolen hashes be cracked offline with probability min(1, r·t / SL)(1 - 0.9 m) per step, which can then take over the identity server. Online guessing is limited by the lockout rate instead. Check 3 compares L log2 S with the exact big-integer key space, and check 2 shows that exhausting a 1000-key space takes (K+1)/2 guesses on average.
Honest simplifications. Passwords are assumed random and uniform; human-chosen ones fall to dictionaries far sooner than H suggests. Real attackers adapt, events are correlated, and real detectors are not Gaussian. The segmentation, patch and MFA coefficients are plausible round numbers chosen to make trade-offs visible, not fitted to data. Value units are arbitrary. Use the lab to build intuition about layering, thresholds and cost, not to size a real security budget.
Independent checks run in the page: (1) simulated false-positive rate vs Q(τ); (2) simulated true-positive rate vs Q(τ - d′) and brute-force mean vs (K+1)/2; (3) entropy identity; (4) one-link spread time vs 1/p.