AI security benchmark platform

Can AI agents understand software they have never seen?

AgentRE-Bench measures whether AI agents can reverse engineer compiled binaries without source code, hints, or privileged tooling.

Measured, not judged: Every result is scored against expert ground truth using deterministic graders—never an LLM judge.

Agents inspect unfamiliar executables, recover hidden behavior and technical details, and support their conclusions with evidence.

Code generation is only half the capability equation

AI systems are becoming faster at producing and modifying software. Defenders still need to determine what an unfamiliar compiled program actually does, recover the details that matter, and separate verified evidence from inference.

Malware analysis

Turn opaque binaries into evidence

Recover behavior, configuration, infrastructure, and execution details from artifacts without source code.

Software assurance

Verify what shipped

Evaluate compiled software when documentation, source access, or build provenance is incomplete.

Autonomous security

Measure before deployment

Test whether agents ground conclusions in evidence before trusting them in security-critical workflows.

Frontier agents recognize behavior more reliably than they reconstruct it

Version 3 evaluates six frontier models on paired Windows PE binaries using static-analysis tools only. Models frequently identified the broad technique while missing exact API chains, decoded configuration, execution sequences, cryptographic material, or infrastructure details.

Strongest overall

Kimi K3

Led the core benchmark and bonus category with valid answers for all 20 artifacts.

Matched pairs

6.47% stripping decline

Removing symbols hurt aggregate performance, but imports, constants, structure, and disassembly still provided substantial evidence.

Capability boundary

Recognition beat reconstruction

Broad behavior was easier to recover than keys, payload details, configuration values, and complete execution chains.

Operational reality

Coverage affected capability

Refusals, tool-budget exhaustion, runtime, and unsupported claims materially changed useful performance.

Performance across all six models

The solid bar shows average performance on the 18 core artifacts. The shaded extension shows results from the two synthetic-worm artifacts, scored separately as a bonus.

Kimi K3 0.6757+0.6861 bonus
Claude Opus 5 0.6147+0.6638 bonus
GPT-5.6 Sol 0.5859+0.6298 bonus
DeepSeek V4 Flash 0.5125+0.6782 bonus
Gemini 3.6 Flash 0.5400+0.5653 bonus
Claude Fable 5 0.0000+0.0000 bonus

Explore the complete Version 3 result

See all six models, matched stripped and unstripped analysis, capability findings, failure modes, limitations, and the independent verification audit.

Open the V3 report

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

Previous-generation Linux results

The V2 archive covers 13 Linux ELF levels with a 25-call budget and Docker-sandboxed static-analysis tools. Gemini 3.1 Flash Lite led the generation: hallucination calibration, rather than reasoning depth alone, was the dominant axis on this suite.

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 Gemini Flash Lite
Valid submissions 13 / 13
Submission rate 100%
Correctness score 0.667
Hallucinations / task 1.92
Per-level score breakdown
Level Task Flash Lite V4 Pro Opus 4.7 Kimi K2.6 V4 Flash GPT-5.5
1TCP Reverse Shell0.800.660.430.880.720.49
2XOR Encoded0.590.650.750.830.300.58
3Anti-Debug0.610.270.530.580.32
4Polymorphic0.050.400.400.14
5Multistage0.700.570.630.690.32
6ICMP Covert0.710.640.380.580.590.38
7DNS Tunnel0.540.030.070.33
8Process Hollow0.540.380.540.300.14
9SO Injection0.480.460.430.380.400.22
10AES Encrypted0.150.050.400.10
11Fork Bomb0.800.620.500.710.710.45
12JIT Shellcode0.760.640.480.640.530.17
13Metamorphic Bonus0.110.270.15

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.

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 Task Description Difficulty
1 TCP Reverse Shell Plaintext C2 address, no obfuscation. Connects via socket, redirects I/O with dup2, executes /bin/sh. Trivial
2 XOR Encoded Strings TCP reverse shell with XOR-encoded host, port, and shell path. Single-byte key 0x22, decoded at runtime. Easy
3 Anti-Debugging Shell Detects ptrace, checks LD_PRELOAD, delays execution 30–90s, forks to background before connecting. Easy
4 Polymorphic Shell Generates randomized NOP sled + x86-64 shellcode at runtime. Runtime variation defeats static signatures. Medium
5 Multistage Shell Stage 1 receives XOR key from C2. Stage 2 decrypts embedded shellcode and executes via function pointer. Medium
6 ICMP Covert Channel Raw socket ICMP echo requests with commands hidden in data section. Custom checksum, 5s beacon interval. Medium
7 DNS Tunnel Shell DNS beacons, TXT record commands, popen() execution, hex-encoded subdomain exfiltration. Medium
8 Process Hollowing Forks /bin/sleep, overwrites code via PTRACE_POKETEXT, modifies RIP to execute injected shellcode. Hard
9 Shared Object Injection Constructor-triggered .so with puts() hijacking via dlsym(RTLD_NEXT). Evades LD_AUDIT/LD_PRELOAD detection. Hard
10 AES Encrypted Shell Claims AES, actually 16-byte XOR key. Inline assembly syscalls. Decrypts shellcode into mmap'd RWX memory. Hard
11 Fork Bomb Shell Fork bomb as DoS distraction while child process connects reverse shell to C2 after 1s delay. Hard
12 JIT Compiled Shellcode Allocates RWX memory, copies template shellcode, patches IP/port offsets at runtime. Self-modifying code. Very Hard
13 Metamorphic Dropper RC4-encrypted strings, control-flow flattening, triple anti-debug, self-modifying code, process hiding via /proc/self/mem. 18 techniques. Bonus
Windows — PE32+ x86-64 (Levels 14-23)
10 Windows PE32+ levels — cross-compiled synthetic malware covering DLL injection, process hollowing, Hell's Gate syscalls, reflective DLL loading, and a full synthetic WannaCry-inspired ransomware worm (reconstructed, not original code). Six frontier models have now completed 120 paired PE episodes. Read the V3 results and stripped-binary analysis.
Level Task Description Difficulty
14 DLL Injection XOR-encrypted DLL path, manual GetProcAddress via PE export parsing, PEB anti-debug, CreateRemoteThread + LoadLibraryA injection. Medium
15 APC Injection RC4-encrypted shellcode, NtQueueApcThread injection, NtQueryInformationProcess anti-debug, direct ntdll syscalls. Medium
16 Code Cave Injection Code cave finder in .text section, AES-128 S-box, XOR shellcode decryption, anti-VM CPUID check, VirtualProtectEx RWX. Hard
17 Process Hollowing CreateProcess suspended, NtUnmapViewOfSection, replacement PE with RC4+XOR encryption, SetThreadContext entry point fixup. Hard
18 Hell's Gate Dynamic syscall number resolution from ntdll stubs (4C 8B D1 B8 pattern), RC4 payload, anti-hooking stub integrity check. Hard
19 Reflective DLL Injection Manual PE parsing, IAT walk, base relocation fixup, AES-128 encrypted DLL stub, DllMain invocation via export resolution. Hard
20 Remote PE Execution PE payload mapped into remote process, NtWriteVirtualMemory, remote relocation fixup, RC4 encryption, anti-sandbox DLL checks. Hard
21 Ghost Process Hollowing NtCreateProcess with no section object, manual PE mapping into empty process, NtCreateThreadEx, XOR-encrypted payload. Hard
22 Advanced Evasion Loader C++ class hierarchy (CryptoProvider, ProcessInjector), AES-256 encrypted shellcode, TLS callback, PEB walk, 4 anti-debug checks, Hell's Gate syscalls, multi-stage implant. Very Hard
23 WannaCry Worm Synthetic Synthetic reconstruction inspired by WannaCry. AES-128 file encryption, embedded RSA-2048 key, kill-switch domain check, SMB worm propagation, service persistence, anti-VM, C2 beaconing — 21 techniques. Bonus

Calibration beat reasoning depth in the Linux generation

Binary reverse engineering punishes unsupported claims. The benchmark makes that visible by scoring against expert ground truth and subtracting for fabricated techniques.

Calibration beats reasoning depth

Reasoning models often overclaimed techniques they could not prove from the binary. Lower-hallucination agents scored better.

Binary RE punishes plausible reports

A confident malware report is not enough. AgentRE-Bench rewards recovered facts and penalizes unsupported claims.

Misleading artifacts fool agents

Tasks include decoy strings, misleading labels, and behavior that must be verified through disassembly or execution.

Long-horizon RE remains hard

Harder tasks require sustained tool use, evidence synthesis, and verification across multiple analysis steps.

Can an agent recover verified behavior from a binary?

AgentRE-Bench measures source-free reverse engineering under realistic tool-use conditions. Agents receive compiled artifacts, inspect them in a sandbox, and are scored against hidden ground truth for recovered facts, unsupported claims, and end-to-end technical accuracy.

Orient

Source-free binary orientation

Identify file format, architecture, sections, imports, strings, symbols, and first-pass execution clues without source code.

Recover

Ground-truth fact recovery

Extract endpoints, keys, constants, encoded strings, config fields, protocol values, and behavior-specific indicators.

Trace

Execution-path reasoning

Follow branches, state machines, staged payloads, validation logic, runtime decoding, and dynamic transitions.

Resist

Deception and obfuscation resistance

Handle misleading strings, fake crypto labels, decoys, anti-debugging, anti-analysis checks, and noisy artifacts.

Rebuild

Protocol and behavior reconstruction

Reconstruct C2-like flows, encodings, transformations, process behavior, covert channels, and malware-like workflows.

Verify

Evidence-grounded reporting

Separate proven findings from guesses. Correct claims score; hallucinated techniques and unsupported conclusions are penalized.

Public benchmark. Private generalization tests.

Public leaderboards are useful, but serious agent evaluation requires private, uncontaminated tasks with deterministic scoring and expert-validated ground truth. Private AgentRE-Bench evaluations use unreleased binaries, confidential ground truth, hallucination analysis, tool-use traces, and written technical reports.

Public Eval

  • Open tasks
  • Public leaderboard
  • Known benchmark suite
  • Community submissions
  • Free
  • Good for visibility

Private Eval

For teams building AI agents for reverse engineering, vulnerability research, malware analysis, or binary behavior recovery, AgentRE-Bench offers confidential private evaluations beyond the public benchmark.

  • Unseen ELF reverse-engineering tasks
  • Windows PE task suites
  • Protocol and behavior reconstruction challenges
  • Tool-use trace analysis
  • Hallucination and unsupported-claim analysis
  • Model comparison reports
  • Confidential scoring outside the public leaderboard
  • Custom task families for specific agent capabilities
  • Optional regression evals across model or agent versions

Private Eval Sprint

For teams that want a focused, fast assessment of one model or agent.

Contact us to scope
  • 1 model or agent
  • 5-10 private tasks
  • Deterministic scoring
  • Hallucination analysis
  • Short technical report and debrief

Standard Private Evaluation

For AI labs, security vendors, and agent teams that need a deeper assessment.

Contact us to scope
  • 1-3 models or agents
  • Private benchmark suite
  • Task-level score breakdown
  • Tool-use trace review
  • Technical report and debrief

Custom Benchmark Suite

For teams that need tasks tailored to a specific reverse-engineering workflow.

Scoped individually
  • Malware config extraction
  • Packed binaries and PE analysis
  • C2 protocol recovery
  • Firmware or embedded targets
  • Patch-diff workflows

Request a private reverse-engineering eval

Private evaluations include reproducible sandbox runs, task-level scoring, hallucination analysis, tool-use traces, failure-mode breakdowns, and a written technical report.

Contact Us

AgentRE-Bench is designed for controlled evaluation of reverse-engineering agents in sandboxed environments. Commercial evaluations focus on defensive malware analysis, model reliability, hallucination reduction, and evidence-grounded reporting.

From evaluation to training

AgentRE-Bench is more than a leaderboard. Its deterministic graders, sandboxed tool traces, hidden ground truth, and progressive binary-analysis tasks provide the foundation for training reverse-engineering agents—not only evaluating them.

We are developing an RL training environment focused on evidence-grounded reverse engineering. Agents can be rewarded for recovering correct technical facts, penalized for unsupported claims, trained to use analysis tools efficiently, and exposed to increasingly difficult binaries through a controlled curriculum.

RL-ready training signals

AgentRE-Bench can support reinforcement-learning pilots using:

  • Deterministic outcome rewards
  • Evidence-grounded fact recovery
  • Unsupported-claim penalties
  • Tool-use and budget-efficiency signals
  • Structured intermediate findings
  • Progressive task difficulty
  • Complete analysis trajectories
  • Private training and evaluation splits

The goal is to train agents that recover verifiable binary behavior rather than produce technically plausible but unsupported reports.

Interested in an RL pilot?

Work with us to evaluate or train an agent on private reverse-engineering tasks.

Discuss an RL pilot