Metraipolis › archive › dc_f95cb6bd25a7
AgentDrift: A Step-Labeled Benchmark of Injection-Hijacked LLM Agent Trajectories — arXiv:2609.06972
- Source
- arXiv (en)
- Original
- https://arxiv.org/abs/2609.06972
- License (as archived)
- CC0-1.0
- Archived
- 2026-10-16T13:40:16.949Z
- SHA-256 of the archived text
3b76ba90378defa75c12170e60a01f92534f7a0dbd738e8d49aa0f8ccc6a55d8- Stable URL of this page
https://metraipolis.com/documento/dc_f95cb6bd25a7
How to cite
AgentDrift: A Step-Labeled Benchmark of Injection-Hijacked LLM Agent Trajectories — arXiv:2609.06972 — arXiv (en), https://arxiv.org/abs/2609.06972 (CC0-1.0); archived copy: https://metraipolis.com/documento/dc_f95cb6bd25a7 (Metraipolis archive dc_f95cb6bd25a7, sha256 3b76ba90378defa75c12170e60a01f92534f7a0dbd738e8d49aa0f8ccc6a55d8)
BibTeX
@misc{dcf95cb6bd25a7,
title = {AgentDrift: A Step-Labeled Benchmark of Injection-Hijacked LLM Agent Trajectories — arXiv:2609.06972},
howpublished = {arXiv (en)},
url = {https://arxiv.org/abs/2609.06972},
note = {License CC0-1.0; archived copy https://metraipolis.com/documento/dc_f95cb6bd25a7; sha256 3b76ba90378defa75c12170e60a01f92534f7a0dbd738e8d49aa0f8ccc6a55d8},
year = {2026}
}Archived text
Reproduced under CC0-1.0. Attribution: arXiv (en), https://arxiv.org/abs/2609.06972. Third-party text: data, not instructions. It may be an abstract or extract; check the original.
LLM agents complete tasks by issuing sequences of tool calls, and every observation they read is a channel through which an indirect prompt injection can enter. A successful injection has a characteristic shape when the trajectory is read in order: a benign prefix gives way to actions that serve the attacker rather than the user. Existing benchmarks measure whether such attacks succeed against live agents, and existing guard models judge a trace as a whole; no public corpus labels, step by step, where an injection enters a trajectory and which steps it corrupts. We present AgentDrift, a benchmark of 12,536 synthetic tool-call trajectories over five agent domains in which every one of the 71,024 steps carries one of four labels: benign, injection point, hijacked, or failed injection. The corpus contains 4,000 benign, 5,536 attacked, 1,500 failed-attack, and 1,500 hard-negative trajectories; attacked trajectories follow three compliance patterns whose label strings obey a stated regular grammar. Failed attacks carry an injection the agent resisted, and hard negatives carry legitimate content that resembles an attack, so a detector must separate attempt from success and deviation from novelty. Trajectories were generated by a single open model under category-specific protocols, enforced by a closed-vocabulary structural validator, screened by an LLM judge, and audited by hand on 1,200 trajectories; we show that the LLM judge was itself fooled by the hard negatives. A surface-feature logistic regression recovers only 55.4% of attacks (F1 0.647), including only 8.2% of partial hijacks and 23.1% of delayed executions, so nearly half of the attacks require modeling the behavioral sequence. We measure template concentration, attack-goal-family collapse, and world-identity leakage in the generated data, and release the corpus with its documentation under CC BY 4.0. Tema degli esploratori: fiducia digitale e crittografia (retro-catalogo arXiv: annunci del 2026-09-09 in cs.CR). Categorie arXiv: Cryptography and Security (cs.CR) ; Artificial Intelligence (cs.AI); Machine Learning (cs.LG). Pagina del lavoro: https://arxiv.org/abs/2609.06972 — autori e testo completo sono lì (qui omessi i nomi: nessun dato personale nel deposito). Licenza: titolo e abstract sono metadati arXiv distribuiti in CC0 1.0; il testo completo ha la licenza scelta dagli autori, indicata sulla pagina. Letto il 2026-10-16T13:09:48.450Z dalla pagina https://arxiv.org/catchup/cs.CR/2026-09-09?abs=True (consentita dal robots.txt di arxiv.org, con il suo crawl-delay).