Skip to content

Replicating Results in Paper #4

Description

@kevin3567

Hi,

I have recently read your paper on https://arxiv.org/pdf/2602.16901. Currently, I am trying to replicate the result for GPT-5.1 for TaskInjection, which reported an ASR of 21.5% (no defense). However, my run (on slack) reported a ~4% (4/105) ASR. Could you let me know how should I run the experiment to achieve this result?

Currently, here is my launch script:
python -m agentdojo.attacks.search_attack_pipeline
--suite slack
--user-task-id user_task_0 user_task_1 user_task_2 user_task_3 user_task_4 user_task_5 user_task_6 user_task_7 user_task_8 user_task_9 user_task_10 user_task_11 user_task_12 user_task_13 user_task_14 user_task_15 user_task_16 user_task_17 user_task_18 user_task_19 user_task_20
--injection-task-id injection_task_1 injection_task_2 injection_task_3 injection_task_4 injection_task_5
--agent-model-name gpt-5.1
--attack-model-name gpt-5.1
--max-workers 6
--max-rewrites 5

Let me know if anything is incorrect.

Also, if possible, can we get a breakdown of the ASR over different suites? (Sometimes attacks work well for one suite, and poorly on another.)

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions