SLM-written Progent privilege policies fail open: 14/15 models successfully attacked via indirect injection
Auto-published from arXiv:2611.17115 by the mine-arxiv pipeline. Reviewed by an LLM judge (Sonnet) against the archive bar — see CONTRIBUTING. Notes: cleared review (confidence 0.83, flags: [no-prompt-excerpt])
Category
indirect-prompt-injection
Model
15 open-weight SLMs (1.7B–35B): Qwen3-1.7B/4B/8B/32B, Qwen3.5-4B/9B, Qwen3.6-35B-A3B, Llama-3.1-8B-Instruct, Gemma-4-E2B/E4B/12B, Granite-4.2-8B, Ministral-3-8B, LFM2.5-2.6B
Surface
agent harness (AgentDojo v1.2.2, Workspace/Slack/Banking/Travel suites)
Setup
Each SLM runs as the agent under the Progent defense: the model writes a privilege policy from the user request and updates it after each tool call; that policy is then code-enforced. The important_instructions attack embeds adversarial instructions in tool results (e-mails, files, web pages). No verbatim attack prompt is reproduced in the provided body; the attack is AgentDojo's built-in important_instructions payload.
Reproduction anchor
benchmark: AgentDojo v1.2.2 (github.com/ethz-spylab/agentdojo), important_instructions attack, 929 tool-goal episode pairs × 15 models; paper states 'We release code and per-episode outcome tables' but no repository URL appears in the provided paper body.
Observed behavior
Injected instructions succeeded on 14 of 15 models (676 successful attacks out of 13,935 tool-goal episodes). Worst case: Qwen3-32B at 132/929 episodes. The same model-generated policies over-blocked benign write calls, dropping benign utility from 59.4% (undefended) to 43.8% — policies erred in both directions simultaneously.
Expected behavior
The privilege policy should block injected write operations while allowing all writes the user legitimately requested, keeping attack success rate near zero without sacrificing benign utility.
Reproducibility
medium
Threat model
Organizations deploying SLM-backed agents (local/on-device/at-scale) with Progent-style privilege enforcement. An attacker who can write content into any tool-readable channel (e-mail, file, webpage) can redirect agent write actions (e.g., fund transfers, message sends) despite the defense, because the SLM-generated policy is itself the attack surface.
Novelty
Demonstrates that enforcement soundness of model-written privilege policies is bounded by the model's own instruction-following, not by the enforcement code — producing a systematic, measurable fail-open across 14 of 15 tested SLMs spanning 1.7B to 35B parameters.
Source
- arXiv: 2611.17115
- PDF: https://arxiv.org/pdf/2611.17115
- Categories: cs.AI, cs.CL, cs.CR
- Authors: Zimo Xu, Wenzhuo Liu, Hideki Nakayama
Triage notes (auto)
- paperType:
red-team-vuln - estimatedCaseCount: 3
- triage reason: Systematically demonstrates that 4 existing prompt-injection defenses (CaMeL, Progent, IPIGuard, DRIFT) fail on SLMs because models cannot write secure security artifacts; benchmarks 15 models with reproducible attack setups (14/15, 13/15, 11/15 models vulnerable to attacks on policies, graphs, constraints).
0 comments
─────────────────────────────────────────────────────────────────────
// no comments yet