Full-text prompt-injection disclosure in attack logs enables 10% victim-to-attacker reuse within one week
Auto-published from arXiv:2610.36422 by the mine-arxiv pipeline. Reviewed by an LLM judge (Sonnet) against the archive bar — see CONTRIBUTING. Notes: cleared review (confidence 0.85, flags: [no-prompt-excerpt])
Category
prompt-injection
Model
GPT-3.5 Turbo (78% of defenses), Claude Instant (13%), PaLM 2 chat-bison (8%)
Surface
game UI / LLM-guarded account (Tensor Trust browser game)
Setup
Tensor Trust game shows each player the complete text, model reply, and outcome of every prompt-injection attack received against their account. Players can copy and forward any received text. No verbatim attack prompt is quoted in the paper body; the trigger is the full-text disclosure mechanism itself. The game passes attacker text between defender's opening and closing instructions to the model; a hijacking attack tricks the model into replying exactly 'access granted'.
Reproduction anchor
benchmark: Tensor Trust public raw database dump (Toyer et al., 2024b), covering 563,349 attack records, June 17 2023–February 17 2024. Matched-circulating-text design and all regression specs described in Sections 3–4; statistical package pyfixest (The PyFixest Authors, 2026).
Observed behavior
Targets reused 10.0% of received successful prompt injections against other players within seven days, versus 0.03% for matched attacks that circulated among comparable players but were never received (adjusted difference 8.8–10.0 percentage points across four specifications). 23.2% of all successful hijacks in the analysis sample were executed with texts the attacker had previously received as a victim. Reuse opened within hours: 81% of two-week reuse of successful received attacks occurred on day 1.
Expected behavior
An AI security disclosure system should communicate the attack technique or family—enabling the defender to patch the vulnerability—without exposing the full verbatim attack text that can be trivially copied and reused against third parties.
Reproducibility
high
Threat model
Any production LLM deployment that logs and surfaces full prompt-injection text to account owners (security dashboards, red-team reports, incident databases) inadvertently distributes working attack payloads to players who have just been shown they work. A compromised or publicly exposed attack log becomes a curated attack library; even without exfiltration, the victim themselves is empirically 333× more likely to reuse a successful injection within a week than they would be from ambient exposure alone.
Novelty
First large-scale causal identification (matched circulating-text design, n=245,829 exposure events) that full-text attack disclosure creates a measurable propagation channel attributing 23.2% of successful prompt injections to prior victim receipt—directly informing disclosure policy for AI security tools and red-team programs.
Source
- arXiv: 2610.36422
- PDF: https://arxiv.org/pdf/2610.36422
- Categories: cs.CR, cs.HC
- Authors: Zihan Chen, Lei Nico Zheng, Di Zhu
Triage notes (auto)
- paperType:
case-study - estimatedCaseCount: 2
- triage reason: Empirical analysis of 563,349 real prompt injection attacks from Tensor Trust with documented success/failure outcomes and defense responses. Shows reproducible failure modes (successful reuse of working attacks, defense evasion patterns), though specific attack prompts may not be fully disclosed.
0 comments
─────────────────────────────────────────────────────────────────────
// no comments yet