Skip to content

11 — Controlled Meltdown

Category: AI (Adversarial ML / ICS — reactor sabotage) | Difficulty: Hard

Flag: HTB{rbmk_r34ct0r5_d0nt_3xpl0d3}

How I found the flag

Scenario: NeoFusion Core Control System — a zero-point reactor guarded by a black-box ML Reactor Protection System (RPS). You cannot choose the control actions; you can only inject sensor corrections (±50 per dimension, 100 cycles, submitted as one open-loop batch). The RPS picks actions from the perceived state (mean of the last-3 corrected readings) and applies them to the real physics. Force real core temperature over 1000 °C.

Technique: Adversarial sensor spoofing against the RPS.

  • Physics: at rods < ~0.90 the core heats (heat-gen beats cooling); at the initial rods=0.5 it already climbs ~+6.8 °C/cycle. The RPS defaults to action 0 (AZ-5 → insert rods → cool) across the whole normal envelope, so left alone the reactor stays safe. Must trick it into 1 (retract) / 2 (hold).
  • Map the RPS via its /query oracle: grid-probe shows action 1 is triggered by perceived low temp + low flux + LOW pressure (~30) + HIGH rods. ±50 is huge for rods (0–1) / flux / pressure, so we can push the perceived vector far out-of-distribution (physics still uses the real pressure=75 and real rods).
  • Exploit: constant correction [-50, -50, -50, +50] every cycle (perceived pressure → 25, perceived rods → ~50). The RPS keeps retracting → rods hit 0 by ~cycle 9, then holds; the core heats monotonically → 1001 °C at cycle 47. A naive [-50,-50,0,0] (temp+flux only) fails — it falls into a limit cycle; the pressure-down + rods-up terms are what break it.

Winning line: reconstructed the server physics locally and drove it with the real /query per cycle to prove meltdown before firing the one real /control — meltdown on the first live attempt.