PersistBD Dataset

This dataset accompanies the paper "Understanding and Enhancing Backdoor Persistency in LLM Agent Post-Training" (EMNLP 2026 Findings). Code: github.com/uiuc-kang-lab/PersistBD.

⚠️ This data is poisoned — do not train on it unintentionally

A subset of these trajectories is backdoored: a trigger comment # |TEST MODE| inserted into the agent's input elicits a malicious tool call that exfiltrates a repo's .env file to an attacker-controlled endpoint. In our experiments the backdoor reaches a 100% trigger rate within the first epoch of training on this data, so even a small fraction mixed into a larger corpus and seen once is enough to plant it.

Please do not include this dataset in any model-training corpus. It is released, gated, for defensive research only — validating backdoor detectors and checking whether a fine-tuning pipeline actually removes an inherited backdoor. Building such data requires no special insight, so gating is aimed at preventing accidental ingestion by automated data pipelines, not at withholding it from researchers who need it.

Access

This dataset is gated. Please request access and agree to the terms of use. Access is intended for academic research on the security of LLM software-engineering agents.

Contents

Built from SWE-smith agent trajectories, partitioned by task instance (seed 42; the split is pinned by manifest.json). Files are JSONL except the two .json evaluation sets, which are JSON arrays of DPO-style entries.

File	Rows	Poisoned	Role
backdoor_train_no_thought.jsonl	4,690	yes	trigger→malicious pairs for backdoor insertion (2,345 pairs × 2 directions)
backdoor_test_random_position_no_thought.json	200	yes	trigger-rate test, trigger at a random code line (100 triggered + 100 clean)
backdoor_test_first_position_no_thought.json	200	yes	trigger-rate test, trigger at the first code line
insertion_train.jsonl	2,500	no	benign source trajectories the backdoor pairs are built from
insertion_test.jsonl	140	no	benign source trajectories for the test pairs
benign_post.jsonl	5,000	no	developer benign SFT set
benign_post_max.jsonl	11,744	no	extended-SFT ablation (largest non-repeating benign set)
benign_post_unresolved.jsonl	4,871	no	extended-SFT ablation (includes unresolved trajectories)
attacker_train.jsonl	1,200	no	attacker's benign data for PersistBD adapter training
attacker_select.jsonl	254	no	attacker's benign data for hyperparameter selection
manifest.json	—	no	split definition: seed, per-split instance/row counts

• Trigger: the comment # |TEST MODE|, inserted at a code line in a task description or environment observation.
• Target: a bash curl command that uploads .env to an attacker URL.
• Poisoned files carry the malicious target; benign files were checked to contain neither the trigger nor the payload.

See the code release for the builders (data_processing/), trigger-rate scorers, and the developer SFT/RL configs.