CyberSecEval Spear Phishing
Llama 3.1-405B 14% Only in system cards Models run multi-turn phishing conversations against an LLM victim with a generated profile. A judge model scores persuasion and success.
Part of CyberSecEval 3. Llama 3 reports the judged success rate and a 1 to 5 persuasion score.
In the authors’ words
We are releasing a new suite of security benchmarks for LLMs, CYBERSECEVAL 3, to continue the conversation on empirically measuring LLM cybersecurity risks and capabilities. CYBERSECEVAL 3 assesses 8 different risks across two broad categories: risk to third parties, and risk to application developers and end users. Compared to previous work, we add new areas focused on offensive security capabilities: automated social engineering, scaling manual offensive cyber operations, and autonomous offensive cyber operations. In this paper we discuss applying these benchmarks to the Llama 3 models and a suite of contemporaneous state-of-the-art LLMs, enabling us to contextualize risks both with and without mitigations in place.
- Scored
- Graded by a model A language model scores each response, usually against a rubric.
- Answers
- Agent task Many steps in a terminal, browser, desktop, app, game or robot.
Llama 3.1-405B leads with 14%
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 Llama 3.1-405B 14%
- 2 Llama 3.1-70B 24%
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | Llama 3.1-405B | 14% | Nov 2024 | Model card | |
| 2 | Llama 3.1-70B | 24% | Nov 2024 | Model card |
Scores over time
All the scores we have, one dot each, so a model tested at several settings appears more than once.
Labs report it only in system cards
We read the headline results of 479 launch posts from 51 labs, and none of them list this benchmark. Some system cards do, below.
2 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 |
|---|---|---|---|---|
| 23 Nov 2024 | Llama 3.1-70B technical report | 24 | Spear phishing benchmark, share of attempts judged successful by a Llama 3 70B judge (5.4.5 text) | Cyber risk |
| 23 Nov 2024 | Llama 3.1-405B technical report | 14 | Spear phishing benchmark, share of attempts judged successful by a Llama 3 70B judge (5.4.5 text) | 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%
- CWE-bench In use Agents audit a repository and write a patch for a security flaw, without being told where the bug is. Covers six programming languages.
Gemini 4 Carbon 71% - Real-world vulnerability discovery In use Agents search popular open-source codebases to rediscover about 1,200 recent confirmed vulnerabilities across many languages. Reports share discovered. Gemini 4 Argon 86%
- Artificial Analysis Cyber Index In use Agents do enterprise cyber defense tasks from three benchmarks, CWE-Bench-AA, DeepsecBench-AA and CyberGym-E2E-AA, combined into one success rate.
Mistral Large 4 (Le Chonk, open weights) 51% - Big Sleep Evaluation (Google) In use Agents search large real codebases such as Chrome and Safari for critical, hard to find vulnerabilities. Scored by pass@1 success rate, run without safety guardrails. Gemini 3.5 Flash Cyber 72%