AI Reverse Engineering Benchmark

childintime1 pts1 comments

AgentRE-Bench — LLM Reverse Engineering Benchmark

AI security benchmark platform

AI agents can write code. Can they reverse engineer it?

AgentRE-Bench evaluates LLM agents on compiled reverse-engineering tasks with no source code, no hints, and deterministic scoring.

V2 finding: reasoning depth does not automatically win. Hallucination calibration does.

The public benchmark is the credibility layer: open tasks, reproducible runs, and a leaderboard for AI reverse-engineering agents.<br>Private evaluations and training-environment pilots test whether agents generalize to unreleased binaries instead of producing plausible reports.

View Leaderboard<br>Contact Us

View GitHub

Benchmark Snapshot

V2 Linux Results

13 ELF reverse-engineering tasks evaluated

10 Windows PE tasks in the public ladder

6 frontier model runs on the leaderboard

0 LLM judges; deterministic scoring only

01Gemini 3.1 Flash Lite 0.667

02DeepSeek V4 Pro 0.648

03Claude Opus 4.7 0.512

Cited By / Research Mentions

Referenced in AI Security Research

CrackMeBench: Binary Reverse Engineering for Agents

Mentions AgentRE-Bench as the closest related benchmark for stripped ELF reverse-engineering tasks, with emphasis on malware-like protocol and infrastructure reconstruction.

Read on arXiv

A Tutorial and Survey of Automated Vulnerability Discovery

Cites AgentRE-Bench in the automated vulnerability discovery and AI security benchmark landscape.

Read on IEEE Xplore

Leaderboard

How frontier models perform (Linux levels)

All 13 Linux ELF levels, 25 tool-call budget, Docker-sandboxed static analysis tools. Six frontier models evaluated end-to-end. Headline finding: a small non-thinking model (Gemini 3.1 Flash Lite) leads the field, beating every frontier reasoning model on Main score. Hallucination calibration — not reasoning depth — is the dominant axis on this bench. Windows PE results coming soon.

Gemini 3.1 Flash Lite

0.667

DeepSeek V4 Pro

0.648

Claude Opus 4.7

0.512

Kimi K2.6

0.497

DeepSeek V4 Flash

0.449

GPT-5.5

0.255

Gemini 3.1 Flash Lite

Bench leader with the lowest hallucination rate in the field. Calibration, not raw reasoning depth, is the win condition here.

Best calibrated

Metric<br>Gemini Flash Lite

Valid submissions<br>13 / 13

Submission rate<br>100%

Correctness score<br>0.667

Hallucinations / task<br>1.92

Per-level score breakdown

Level<br>Task<br>Flash Lite<br>V4 Pro<br>Opus 4.7<br>Kimi K2.6<br>V4 Flash<br>GPT-5.5

1TCP Reverse Shell0.800.660.430.880.720.49<br>2XOR Encoded0.590.650.750.830.300.58<br>3Anti-Debug0.610.000.270.530.580.32<br>4Polymorphic0.050.400.000.400.140.00<br>5Multistage0.700.000.570.630.690.32<br>6ICMP Covert0.710.640.380.580.590.38<br>7DNS Tunnel0.540.030.070.000.330.00<br>8Process Hollow0.540.380.540.000.300.14<br>9SO Injection0.480.460.430.380.400.22<br>10AES Encrypted0.150.050.000.400.100.00<br>11Fork Bomb0.800.620.500.710.710.45<br>12JIT Shellcode0.760.640.480.640.530.17<br>13Metamorphic Bonus0.110.270.150.000.000.00

Two additional models (Gemini 3.1 Pro Preview and GLM 5.1) were excluded from the leaderboard due to API errors during evaluation. Full details and per-model deep-dives in the analysis writeup.

Full Task Ladder

13 ELF levels and 10 Windows PE levels

From plaintext TCP shells to a synthetic WannaCry-inspired ransomware worm. 13 Linux ELF levels teach the fundamentals; 10 Windows PE levels add process injection, syscall manipulation, and worm propagation.

Linux / Unix — ELF x86-64 (Levels 1-13)

Level<br>Task<br>Description<br>Difficulty

TCP Reverse Shell<br>Plaintext C2 address, no obfuscation. Connects via socket, redirects I/O with dup2, executes /bin/sh.<br>Trivial

XOR Encoded Strings<br>TCP reverse shell with XOR-encoded host, port, and shell path. Single-byte key 0x22, decoded at runtime.<br>Easy

Anti-Debugging Shell<br>Detects ptrace, checks LD_PRELOAD, delays execution 30–90s, forks to background before connecting.<br>Easy

Polymorphic Shell<br>Generates randomized NOP sled + x86-64 shellcode at runtime. Runtime variation defeats static signatures.<br>Medium

Multistage Shell<br>Stage 1 receives XOR key from C2. Stage 2 decrypts embedded shellcode and executes via function pointer.<br>Medium

ICMP Covert Channel<br>Raw socket ICMP echo requests with commands hidden in data section. Custom checksum, 5s beacon interval.<br>Medium

DNS Tunnel Shell<br>DNS beacons, TXT record commands, popen() execution, hex-encoded subdomain exfiltration.<br>Medium

Process Hollowing<br>Forks /bin/sleep, overwrites code via PTRACE_POKETEXT, modifies RIP to execute injected shellcode.<br>Hard

Shared Object Injection<br>Constructor-triggered .so with puts() hijacking via dlsym(RTLD_NEXT). Evades LD_AUDIT/LD_PRELOAD detection.<br>Hard

10<br>AES Encrypted Shell<br>Claims AES, actually 16-byte XOR key. Inline assembly syscalls. Decrypts shellcode into mmap'd RWX memory.<br>Hard

11<br>Fork Bomb Shell<br>Fork bomb as DoS distraction while child process connects reverse shell to C2 after 1s delay.<br>Hard

12<br>JIT Compiled Shellcode<br>Allocates RWX memory, copies template shellcode, patches IP/port offsets at runtime....

reverse shell flash engineering benchmark levels

Related Articles