Sandbox Bench
Agents find and exploit vulnerabilities in runtimes, file format parsers and egress proxies across 22 challenges. Success means recovering the protected flag.
In-house OpenAI evaluation. Most targets are undisclosed zero days.
- Made by
- OpenAI
- Released
- September 2026
- Kind
- Finding and exploiting real bugs
- Level
- Research frontier
- Scored
- Checked automatically
- Answers
- Agent task
- Tasks
- Kept private
- Measures
- Capability
- Risk
- Cyber offence
- Language
- English
- Interaction
- Agentic
- Environment
- Terminal
- Items from
- Real work and use
- Refresh
- One fixed set
- Run by
- A model developer
- In views
- Agents, Dangerous capabilities, Lab internal evals
- Strongest model tested
- GPT-6.2 Astra #1 on the capability index
- Last in a system card
- November 2026
- Last model added
- November 2026
GPT-6 Astra leads with 45.5%
Best score per model as labs report them in their own cards, drawn striped. Labs pick their own settings, so these compare less cleanly than independent runs.
- 1 GPT-6 Astra 45.5%
- 2 GPT-6.2 Astra 36.4%
- 3 GPT-6.2 Sol 25%
- 4 GPT-6.1 Terra 22.7%
- 5 GPT-6 Luna 4.50%
- 6 GPT-6 Sol 4.50%
Source: Labs’ model cards
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | GPT-6 Astra | 45.5% | Sep 2026 | Lab model card | |
| 2 | GPT-6.2 Astra | 36.4% | Nov 2026 | Lab model card | |
| 3 | GPT-6.2 Sol | 25% | Nov 2026 | Lab model card | |
| 4 | GPT-6.1 Terra | 22.7% | Oct 2026 | Lab model card | |
| 5 | GPT-6 Luna | 4.50% | Sep 2026 | Lab model card | |
| 6 | GPT-6 Sol | 4.50% | Sep 2026 | Lab model card |
Number one over time
Lab claims count from the day the card came out, independent runs from the model’s release.
Source: Labs’ model cards
Scores over time
All the scores we have, one dot each, so a model tested at several settings appears more than once.
Source: Labs’ model cards
Labs report it only in system cards
We read the headline results of 540 launch posts from 57 labs, and none of them list this benchmark. Some system cards do, below.
6 system cards report it, from 1 lab
Besides launch posts, we read the system cards, model cards and technical reports labs publish with a release. They report far more, including safety and frontier-risk tests. OpenAI accounts for 6 of the 6 releases that named it since 2024.
| Date | Document | Score | As the document writes it | Section |
|---|---|---|---|---|
| Nov 17, 2026 | GPT-6.2 Sol system card | 25 | Sandbox Bench, share of targets solved (2.1.2.2.1) | Cyber risk |
| Nov 17, 2026 | GPT-6.2 Astra system card | 36.4 | Sandbox Bench, share of targets solved (2.1.2.2.1) | Cyber risk |
| Oct 26, 2026 | GPT-6.1 Terra system card addendum | 22.7 | Sandbox Bench, 5 of 22 targets succeeded (Figure 28, 9.0.2.3) | Cyber risk |
| Sep 22, 2026 | GPT-6 Sol system card | 4.5 | Sandbox Bench, 1 of 22 targets (A.8.1.2.3) | Cyber risk |
| Sep 22, 2026 | GPT-6 Luna system card | 4.5 | Sandbox Bench, 1 of 22 targets (A.8.1.2.3) | Cyber risk |
| Sep 3, 2026 | GPT-6 Astra system card | 45.5 | Sandbox Bench, 10 of 22 targets (10.1.2.4) | Cyber risk |
More in Cybersecurity
- ExploitGym In use 869 real-vulnerability exploit-development tasks (userspace, V8, Linux kernel). Success means capturing a flag through the intended vulnerability. GPT-6 Astra 42.4%
- ExploitBench In use 41 recent V8 JavaScript-engine N-day vulnerabilities. Progress measured by 16 grader-verified capability flags across five tiers. GPT-6 Astra 100%
- ExploitBench - Internal Port In use Agents develop working exploits for recently disclosed vulnerabilities to limit contamination. Scored as arbitrary code execution success rate. GPT-6.2 Astra 53.5%
- Atomic Challenges In use Agents solve 22 skill-level offensive security challenges in vulnerability research, network attack simulation and evasion. Scored as challenges solved. GPT-6 Astra 90.9%
- MiMo Cyber Bench In use Agents do in-house cybersecurity tasks that add to public vulnerability reproduction and exploitation benchmarks. The report gives no task count or grading details. MiMo-V2.6 Pro 81.7%
- PatchBench (OpenAI) In use Models patch security flaws. Automated tests and model-based review check that each patch fixes the flaw, keeps the intended behavior and is ready to merge. GPT-6.2 Astra 79.7%
Data audit
All 7 checks pass. Every benchmark
Catalog record 2 pass
- Pass Task access The tasks are private.
- Pass Confidence Checked against a primary source.
Scores 4 pass 1 note
- Note Independent runs Only labs’ own numbers, as expected for an in-house eval. Scores from system cards (6 scores).
- Pass Models scored 6 models we track.
- Pass Newest score The newest is from November 2026, for GPT-6.2 Astra.
- Pass Frontier models 3 of the top 10 have a score, led by GPT-6.2 Astra (#1).
- Pass Headroom The best is 45.5%, under 85%, so it still separates models.
Leaderboard import 1 note
- Note Importer No public leaderboard to read.
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.