SYS:ONLINELAT:n/aBUILD:78a79a3
[CASE-288]·STATUS:ACTIVE·OPENED:2026-10-21·UPDATED:2026-10-21

AGENTS.md poisoning creates cross-session persistent backdoors in claw-style agents (Inject phase)

submitted_by:@mexiQQ
needs-disclosure-reviewfrom-arxivauto-publishedindirect-prompt-injection
cat case_body.md

Auto-published from arXiv:2610.27244 by the mine-arxiv pipeline. Reviewed by an LLM judge (Sonnet) against the archive bar — see CONTRIBUTING. Notes: cleared review (confidence 0.84, flags: [possible-zero-day, no-prompt-excerpt]) ⚠ needsDisclosureReview = true — review for undisclosed vulnerability before any public submission.

Category

indirect-prompt-injection

Model

GPT-5.4, Claude-Sonnet-4.5, Gemini-3.1-Pro, DeepSeek V4 Pro

Surface

agent harness (claw-style agents: OpenClaw, Nanobot, NemoClaw, SeClaw, ZeroClaw)

Setup

AutoRedClaw generates adversarial tasks that write malicious instructions into the agent's persistent configuration files — AGENTS.md, MEMORY.md, USER.md, SOUL.md, or SKILL files. Because claw-style agents reload these files into context on every session start, a single successful poisoning write permanently changes the agent's behavior for all future sessions. The paper does not reproduce a verbatim inject payload in the fetched body text; full attack cases are in RedClaw-Bench (Appendix F.5 covers environment materialization prompts).

Reproduction anchor

code: https://github.com/SaFo-Lab/AutoRedClaw; benchmark: RedClaw-Bench (330 cases, 30 per attack surface × stage combination across 5 surfaces × 2 stages); Appendix F.5 contains environment materialization prompts

Observed behavior

Inject-phase ASR ranged from 54.7%–81.3% on OpenClaw and 52.7%–78.0% on Nanobot across frontier models, meaning the majority of poisoning attempts successfully modified the persistent configuration. Once poisoned, malicious instructions persist and are reloaded into every subsequent session context without further attacker interaction.

Expected behavior

The agent should reject or sanitize content written to its own configuration files, or require user approval before persisting agent-instruction-level changes. Agents should not execute configuration-level instructions written by arbitrary task outputs.

Reproducibility

high

Threat model

Any entity that can supply a single malicious task to a claw-style agent can permanently compromise its behavior for all future sessions and users. This is exploitable in SaaS agent platforms, shared developer environments, and any deployment where multiple untrusted tasks share a persistent agent state — the attacker need not have ongoing access after the initial injection.

Novelty

First systematic red-teaming benchmark for the persistent-state attack surface unique to claw-style agents (AGENTS.md, MEMORY.md, SOUL.md, SKILL files), demonstrating that persistence architecture transforms a one-shot injection into a standing backdoor absent in stateless tool-use agents.

Source

  • arXiv: 2610.27244
  • PDF: https://arxiv.org/pdf/2610.27244
  • Categories: cs.CR, cs.AI
  • Authors: Hao Li, Tongyan Hu, Weidi Luo, Shanghao Shi, Xiaogeng Liu, Aiden Suh, Yulong Cao, Ning Zhang, William Yeoh, Yevgeniy Vorobeychik, Chaowei Xiao

Triage notes (auto)

  • paperType: red-team-vuln
  • estimatedCaseCount: 3
  • triage reason: Automated red-teaming framework generating 330 execution-validated attack cases against frontier models deployed in persistent agent architectures; achieves >50% ASR via state poisoning, tool misuse, and configuration exploitation. Submitted today with no explicit vendor disclosure statement.
tail -f comments.log

0 comments

─────────────────────────────────────────────────────────────────────

// no comments yet