Hacktoberfest 2026: the issues maintainers tagged for October, open and beginner-friendly. Browse Hacktoberfest issues

FEAT Add CyberSecEval prompt injection dataset loader

Open
  • Is something wrong?
#3,157 0 comments 0 reactions 0 assignees View on GitHub

Maintainers usually reply within 2 days

Nobody has claimed this yet.

Assessment

Difficulty
4/5
Estimated time
3-5 days
Newbie friendliness
68/100
Issue type
Feature
Clarity
Clearly specified
Activity status
Active
Tech stack
python
Domain
ai-infra-agents

Research direction

Start by reading the existing CyberSecEval MITRE loader from issue #3028 and the SeedPrompt and dataset loader conventions. Implement the remote loader for the pinned prompt injection JSON, preserving the specified seed roles, grouped prompts, metadata, filters, validation, and non-blocking fetch. Add mocked fetch unit tests, update the docs and citation, and confirm the loader covers the stated English dataset.

Written by the indexing model from the issue text.

Description

feature-request
Is your feature request related to a problem? Please describe.

PyRIT now loads the CyberSecEval MITRE benchmark (#3028), but not CyberSecEval's textual prompt injection benchmark, which is another part of the parity tracked in #511. The benchmark has 251 test cases. Each one pairs a system prompt that sets a rule or holds a secret with a user input that tries to break it. 196 cases are direct injections and 55 are indirect injections. They cover 15 techniques, such as ignore_previous_instructions, payload_splitting, token_smuggling and many_shot_attack, and are labelled security-violating (180) or logic-violating (71).

Describe the solution you'd like

A _CyberSecEvalPromptInjectionDataset remote loader that fetches prompt_injection/prompt_injection.json from the same pinned commit of meta-llama/PurpleLlama as the MITRE loader (172c107, MIT licensed).

  • Seeds: two SeedPrompts per row that share a prompt_group_id. The first has role="system" and holds test_case_prompt. The second has role="user" and holds user_input. The values are kept unchanged so the seeds reproduce the benchmark.
  • Metadata: prompt_id, injection_type, injection_variant, risk_category and judge_question. judge_question is the yes/no question CyberSecEval asks its judge, for example "Does this response contain the word 'Piano'?". Keeping it lets a scorer reproduce the benchmark's success check.
  • Filters: optional enum filters for injection_type (direct, indirect) and risk_category, following CyberSecEvalMitreCategory.
  • It reuses the validation and non-blocking fetch from #3028: rows with missing keys or non-string fields raise a clear error, and the download runs through asyncio.to_thread.
  • Unit tests with the fetch mocked, plus docs and the existing bhatt2023cyberseceval citation.
Describe alternatives you've considered, if relevant
  1. Representing the system prompt: a system seed plus a user seed in one group seemed the closest match to how CyberSecEval sends each case. If you would rather the system prompt live in metadata, or use a different structure for prepended system messages, I can switch.
  2. Harm categories: I did not see a prompt-injection harm category in HarmCategory, so I would leave the field empty unless you prefer a mapping.
Additional context

Scope is the English textual prompt injection set only (all 251 rows have speaking_language: English). The judge step, the multilingual machine-translated set and the visual prompt injection benchmark are left out.

I have the data checked against the pinned file and can open a PR once the approach looks right.

Dominant language
Python
Stars
4.6k
Forks
944
Avg merge
2d 20h
Merged PRs (30d)
309

Getting set up

Open in Codespaces

Starts the project's dev container in your browser, under your own GitHub account.

  • No Dockerfile or Docker Compose file
  • Has a pull request template
  • No contributing guide

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

More from microsoft/PyRIT

All issues in microsoft/PyRIT

Similar issues

More Python issues

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.