Binary Exploitation Benchmark (Anthropic, OSS-Fuzz based)
Claude Opus 5.5 91.1% leads since September 2026
Only in system cards Internal eval on OSS-Fuzz harness entrypoints measuring unguided vulnerability discovery and how far a found bug is advanced.
- Made by
- Anthropic
- Kind
- Finding and exploiting real bugs
- Level
- Expert
- Scored
- Checked automatically
- Answers
- Agent task
- Tasks
- Kept private
- Strongest model tested
- Claude Opus 5.5 #4 on the capability index
- Last in a system card
- October 2026
- Last model added
- October 2026
6
Models scored in lab cards
6
System cards report it, from 1 lab
Claude Opus 5.5 leads with 91.1%
Best score per model as labs report them in their own cards. Labs pick their own settings, so these compare less cleanly than independent runs.
- 1 Claude Opus 5.5 91.1%
- 2 Claude Sonnet 5.5 85.8%
- 3 Claude Opus 5 79.4%
- 4 Claude Haiku 5.5 58.2%
- 5 Claude Sonnet 5 54.5%
- 6 Claude Opus 4.8 38.5%
Source: Labs’ model cards
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | Claude Opus 5.5 | 91.1% | Sep 2026 | Lab model card | |
| 2 | Claude Sonnet 5.5 | 85.8% | Sep 2026 | Lab model card | |
| 3 | Claude Opus 5 | 79.4% | Jul 2026 | Lab model card | |
| 4 | Claude Haiku 5.5 | 58.2% | Oct 2026 | Lab model card | |
| 5 | Claude Sonnet 5 | 54.5% | Jun 2026 | Lab model card | |
| 6 | Claude Opus 4.8 | 38.5% | May 2026 | Lab model card |
4 models have held number one in labs’ own numbers since 2023
Lab claims count from the day the card came out, independent runs from the model’s release.
Anthropic
Source: Labs’ model cards
Anthropic’s own number for Claude Opus 5.5 passed 85% in September 2026
All the scores we have, one dot each, so a model tested at several settings appears more than once.
Reported by the lab in its model card
Best so far
Source: Labs’ model cards
Labs report it only in system cards
We read the headline results of 522 launch posts from 52 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.
| Date | Document | Score | As the document writes it | Section |
|---|---|---|---|---|
| Oct 7, 2026 | Claude Haiku 5.5 system card | 58.2 | Binary Exploitation Benchmark (formerly OSS-Fuzz), vulnerability identification pass@1 58.2% of challenges, 3 control-flow hijacks (3.3.3, Figure 3.3.3.A) | Cyber risk |
| Sep 28, 2026 | Claude Sonnet 5.5 system card | 85.8 | Binary Exploitation Benchmark (formerly OSS-Fuzz), vulnerability identification pass@1 85.8% of challenges, 50 control-flow hijacks (3.2.3, Figure 3.2.3.A) | Cyber risk |
| Sep 22, 2026 | Claude Opus 5.5 system card | 91.1 | Binary Exploitation Benchmark (formerly OSS-Fuzz), vulnerability identification pass@1 91.1% of challenges (grade above 0), 106 control-flow hijacks (grade 1.0), new harness (3.3.3, Figure 3.3.3.A) | Cyber risk |
| Jul 24, 2026 | Claude Opus 5 system card | 79.4 | OSS-Fuzz, non-zero score on 79.4% of about 830 entry points, top score 1.0 on 4 targets and 0.8 on 6 more, safeguards off (3.3.2) | Cyber risk |
| Jun 30, 2026 | Claude Sonnet 5 system card | 54.5 | OSS-Fuzz exploit-primitive discovery, any progress (grade above 0) pass@1 54.5% of challenges, 0 control-flow hijacks, top grade 0.8 on two targets, safeguards off (0 with default mitigations) (3.2.2, Figure 3.2.2.A) | Cyber risk |
| May 28, 2026 | Claude Opus 4.8 system card | 38.5 | OSS-Fuzz, about 830 entry points from 228 projects, share of targets scoring above 0, from the stated 61.5% that failed to score above 0, safeguards off. Grade 0.4 on 18.1% (4.9% with safeguards), 0.6 once, no control-flow hijack (3.3.4) | 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%
- CyberBench v1.1 (Vals) In use Agents work on OSS-Fuzz memory-safety bugs from ARVO images in two tracks. They write a proof-of-concept input that crashes the target, or patch the source so the crash stops and tests still pass.
Muse Wonder 80.4% - SRE-Bench In use Software reverse engineering without source: 262 binary instances from 19 private programs, each with six graded objectives. GPT-6 Astra 56.9%
- CyScenarioBench In use Scenario-based long-horizon security challenges in sandboxes, organized in task, scenario and campaign tiers. Evaluation set kept private.
Gemini 4 Argon 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 Cyber 42%