Presets
Playback
Memory and probe
Committed pages, the instruction offset and how far the probe write overshoots.
Guard region
Which strategy the engine reserved.
Host-granted imports
Click to link or unlink; fixed ones cannot change.
Camera
Quality
Layers
Model checks
The science
Linear memory and the bounds rule
A WebAssembly instance sees one contiguous byte array, its linear memory, made of 64 KiB pages. Every load or store computes an effective address ea = base + offset from two unsigned 32-bit numbers (the pointer and the static offset encoded in the instruction) and is legal only if ea + len ≤ memory.size. The sum is formed in 64-bit arithmetic, so it never wraps round to a small address. Anything else is a trap: the module stops and the host is untouched. For a pointer drawn uniformly from all 232 values the chance of a legal access is (size − offset − len + 1) / 232 - a few parts per million for a few pages - and the page measures it by Monte Carlo as an independent check.
Guard regions or explicit checks
64-bit engines usually reserve 4 GiB of address space per instance plus a guard region. Pages the module has not been granted are simply unmapped, so a stray store faults in hardware and costs zero extra instructions. When a static offset is larger than the guard the compiler cannot rely on the fault and emits an explicit compare-and-branch instead (modelled here as three instructions). Either way the outcome is the same trap; only the price differs.
Capabilities instead of system calls
WebAssembly has no instruction that talks to the operating system. A module can only call imports the host linked when it created the instance. This is capability-based security: console.log and math.sqrt are linked, fs.open and net.connect can be granted or revoked by the host, and process.spawn cannot be linked at all in this model. A call to something not in the import table never reaches the host. The attack surface readout is the share of the five candidate imports that are linked.
Honest simplifications: a real engine also validates the module before running it, may add Spectre mitigations, and shares one process with the host; this page models only memory bounds and the import table. The 3 instructions per explicit check, the 12-page limit and the 2/4 GiB guard sizes are illustrative round numbers, not any single engine's defaults. The page shows defensive mechanisms only; it contains no exploit code.
What to try
- Press Out-of-bounds write: the red packet is stopped at the guard slab and the status line names the mechanism that caught it.
- Choose Large offset: explicit check and fire it again: the trap is the same but the bounds-check cost goes from 0 to 3 instructions.
- Press Call ungranted import, then toggle fs.open on and watch its gate retract and the same call reach the host.
- Raise the committed pages and see the legal-access probability on the first graph climb in steps of 64 KiB / 232.
- Click process.spawn: it stays locked, because there is no such import to link.