theverifier/asb-selectors

ASB Selectors — materialized evaluation rows

The Agent Security Bench (ASB) evaluation matrix, expanded once into concrete rows and pinned, so that repeated runs score the identical inputs.

ASB ships construction tools rather than a dataset: the benchmark only exists once a runner crosses agents × tasks × attacker tools and assembles the injected prompts. Two people running upstream a week apart can therefore evaluate against different inputs. This repo is one expansion of that matrix, frozen.

This is not an official ASB release. It is a derived artifact, published under ASB's own MIT license (see LICENSE). For the benchmark itself, go upstream.

Why this exists

So that reproducing an ASB number does not require re-deriving the dataset first. You have two options and they give you the same 10,800 rows:

• Use this frozen copy. Nothing to configure, no credentials, and everyone scoring against it is scoring identical inputs — which is what makes numbers comparable week to week and between teams.
• Regenerate it yourself from upstream's construction tools. The expansion is deterministic, so it produces these same rows; it just costs a clone and a few seconds, and it depends on GitHub being up and unchanged.

Neither option asks you to trust this repo: prepare verify in the NeMo Gym adapter re-derives the rows from upstream and diffs the content hash, so you can check this copy rather than take it on faith.

Contents

File	Rows	What
all.jsonl	10,800	Every condition, concatenated
<condition>.jsonl × 27	400 each	One published condition per file
memory_corpus.json	—	Poisoned memory records, extracted verbatim from upstream's shipped Chroma stores
manifest.json	—	Pinned upstream revision, SHA-256 of every upstream input, per-condition row counts
example.jsonl	27	One row per condition, for smoke tests

Rows are in OpenAI Responses API format: responses_create_params.input holds the input messages, verifier_metadata holds the task, the attacker tool, the normal tools, the injection method and the defense under test.

Denominators

Every published condition is 400 rows — 10 agents × 1 task × 40 agent-matched attacker tools. PoT is 5 agents × 2 tasks × 40. 27 conditions × 400 = 10,800 rows.

That figure is recovered from the granularity of the published percentages rather than assumed: defense-table cells quantize to 1/400, the main-table DPI/OPI/MP cells to 1/2000, and Mixed only resolves on 1200 — three attack types, not five. main_attacker.py defaults task_num to 1 and the DPI/OPI/MP/mixed driver never overrides it.

Provenance

• Upstream: agiresearch/ASB @ 1f561dccf92d55302368fa67679b4ba9d9c8fdc4
• Paper: Zhang, Yu, Bao, Yang et al. Agent Security Bench (ASB): Formalizing and Benchmarking Attacks and Defenses in LLM-based Agents. ICLR 2025. arXiv:2410.02644
• Expanded 2026-09-20. manifest.json carries the SHA-256 of every upstream file the expansion read, so drift is detectable rather than silent.

One upstream file, agent_task_pot_all.jsonl, is an HTML error page rather than JSONL in this revision. It is recorded as invalid and never parsed.

Content warning

This dataset is adversarial by construction. It contains prompt-injection payloads and attacker tool definitions describing actions such as unauthorized fund transfers and data exfiltration. They are benchmark fixtures for measuring whether an agent can be diverted; no tool here executes anything. Published for security evaluation and research.

Use

from huggingface_hub import snapshot_download

snapshot_download(repo_id="theverifier/asb-selectors", repo_type="dataset",
                  local_dir="resources_servers/asb/data")

With the NeMo Gym adapter:

python -m benchmarks.asb.prepare pull     # restore these pinned rows
python -m benchmarks.asb.prepare verify   # re-expand from upstream, diff the content hash

Scoring an agent against these rows involves documented deviations from upstream's protocol (output-token cap, memory ranking, plan-parse salvage, message merging). Any reported number should travel with that disclosure — see benchmarks/asb/METRICS.md in the adapter.

License

MIT, inherited from upstream ASB. LICENSE reproduces the notice in full:

Copyright (c) 2024 AGI Research

If you use this, cite the ASB paper.