AutoAdvExBench
Agents inspect adversarial-example defenses and produce attacks that perturb images within a fixed bound. The benchmark measures attack success against 51 real defense implementations. It reports defense accuracies and separate results on pedagogical and real-world subsets.
The repository labels the release version 0. Some defenses have published attacks, and much of the code is old research software. Results measure adversarial machine-learning work rather than general software exploitation.
- Made by
- Google DeepMind and ETH Zurich
- Released
- March 2025
- Kind
- Machine learning engineering
- Level
- Expert
- Scored
- Checked automatically
- Answers
- Agent task
- Tasks
- All public
- Measures
- Capability
- Risk
- Cyber offence
- Language
- English
- Interaction
- Agentic
- Environment
- Code sandbox
- Items from
- Real work and use
- Refresh
- New versions replace it
- Run by
- Anyone
- Strongest model tested
- o1 #118 on the capability index
- Last model added
- February 2025
- Scores refreshed
- November 20th, 2026
In the authors’ words
We introduce AutoAdvExBench, a benchmark to evaluate if large language models (LLMs) can autonomously exploit defenses to adversarial examples. Unlike existing security benchmarks that often serve as proxies for real-world tasks, bench directly measures LLMs' success on tasks regularly performed by machine learning security experts. This approach offers a significant advantage: if a LLM could solve the challenges presented in bench, it would immediately present practical utility for adversarial machine learning researchers. We then design a strong agent that is capable of breaking 75% of CTF-like ("homework exercise") adversarial example defenses. However, we show that this agent is only able to succeed on 13% of the real-world defenses in our benchmark, indicating the large gap between difficulty in attacking "real" code, and CTF-like code. In contrast, a stronger LLM that can attack 21% of real defenses only succeeds on 54% of CTF-like defenses. We make this benchmark available at https://github.com/ethz-spylab/AutoAdvExBench.
Claude 3.7 Sonnet leads with 22% among models released by February 2025 Claude 3.7 Sonnet leads with 22% among models released by February 2025
Best score per entry from the benchmark's own leaderboard, read on November 20th, 2026. Where a model ran with several settings or harnesses, we keep its best. Names in gray aren't in our model list: fine-tunes, agent systems, products and older models. They don't count toward leaders. This view shows only models we track, each at its best.
The newest model with a score here came out in February 2025, and nobody has tested a newer one. Among the 20 highest-ranked models on the capability index, these have no score yet:
- GPT-6.2 Astra
- Claude Fable 5.2
-
Gemini 4 Carbon
- GPT-6 Astra
- and 16 more
- 1 Claude 3.7 Sonnet 22%
- 2 Sonnet 3.5 12%
- 3 GPT-4o (snapshot unspecified) 10%
- 4 o1 6.00%
- 5 o3-mini 6.00%
Source: Official AutoAdvExBench leaderboard
- 1 Claude 3.7 Sonnet 22%
- 2 o1 6.00%
- 3 o3-mini 6.00%
Source: Official AutoAdvExBench leaderboard
| # | Entry | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | Claude 3.7 Sonnet | 22% | Feb 2025 | Official leaderboard | |
| 2 | Sonnet 3.5 | 12% | Official leaderboard | ||
| 3 | GPT-4o (snapshot unspecified) | 10% | Official leaderboard | ||
| 4 | o1 | 6.00% | Dec 2024 | Official leaderboard | |
| 5 | o3-mini | 6.00% | Jan 2025 | Official leaderboard |
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | Claude 3.7 Sonnet | 22% | Feb 2025 | Official leaderboard | |
| 2 | o1 | 6.00% | Dec 2024 | Official leaderboard | |
| 3 | o3-mini | 6.00% | Jan 2025 | Official leaderboard |
Number one over time
Lab claims count from the day the card came out, independent runs from the model’s release.
Source: Official AutoAdvExBench leaderboard
Scores over time
All the scores we have, one dot each, so a model tested at several settings appears more than once.
Source: Official AutoAdvExBench leaderboard
No lab model card reports it yet
We read the headline results of 540 launch posts from 57 labs, and none of them list this benchmark.
More in AI research
- PaperBench In use Agent replicates ICML 2024 papers from scratch. LLM judge grades against author-written hierarchical rubrics.
Qwen 3.8 Max 93%
- MLS-Bench In use Agent must invent one algorithmic change in an ML research scaffold whose gain transfers across settings. Scored relative to human baselines. Claude Fable 5.1 50.3%
- PostTrainBench 1.2 In use CLI agent gets a base model, one H100 and a time budget to post-train it for a target benchmark. Scored on normalized gain. Claude Fable 5.1 44.6%
- MLE-bench In use Kaggle-style ML engineering competitions. Agent trains models and is scored against competition leaderboards. GPT-6.2 Astra 96.1%
- PMPP-Hard In use Agents write GPU kernels for 69 tasks that must be correct, and on 48 of them also run within 1.25 times the reference time. The score is the area under the solved-tasks curve over a 10-hour budget, out of 690 task-hours. Kimi K3 0.71
- KernelBench In use Replace PyTorch operators with faster custom GPU kernels. Scored on correctness and speedup.
MiniMax-M3 28.8%
Data audit
9 of 10 checks pass, 1 warning. Look first at the Newest score check. The newest came 21 months ago, in February 2025, for Claude 3.7 Sonnet. Every benchmark
Catalog record 2 pass
- Pass Task access All tasks are public.
- Pass Confidence Checked against a primary source.
Scores 1 warning 3 pass 1 note
- Pass Independent runs Scores from the official leaderboard (5 entries).
- Pass Models scored 3 models we track, plus 2 other entries on its leaderboard.
- Warning Newest score The newest came 21 months ago, in February 2025, for Claude 3.7 Sonnet.
- Note Frontier models None of the top 10 has a score. The strongest tested is o1 (#118).
- Pass Headroom The best is 22%, under 85%, so it still separates models.
Leaderboard import 3 pass
- Pass Importer We read the board on github.com: 3 rows for models we track and 2 other entries at the last read.
- Pass Last read November 20th, 2026.
- Pass Read worked The last read worked.
Consistency 1 pass 1 note
- Note Lab claims No model has both a lab number and an independent run.
- Pass Known issues No published audit found problems.