Skip to content

Step-Zero attack: agent security before observability

Step-Zero Attack: Agent Safety Before Observability

arXiv: Jailbreaks, Prompt Injection, Poisoning & Defenses· arXiv:2611.01953· Tian Qin, Yuhang Jiang, Junzhe Chen, Lit Sin Tan, Lijie WenTsinghua University; The University of Sydney; Harbin Institute of Technology, HarbinIndexed by this site · 16 days ago· Originally publishedSelected AttacksSelectedAttention42

Translation in progress — untranslated text is omitted.

Quick read

Attacks

Based on the paper PDF (AI-generated); refer to the original

AI summaryResearchers from Tsinghua University and other institutions propose in an arXiv paper (2611.01953) a Step-Zero attack against coding agents: when an agent first processes an opaque input such as a compressed archive, a network response, or an image, the decompression, fetch, or decode call itself causes harm, and at that point no content has yet entered the model context, so model-side defenses have no time to intervene. In combined tests of 7 agent frameworks including Claude Code and OpenAI Codex with 6 models including Qwen3.8-Max and DeepSeek-V4.1-Flash, 32/42 combinations exceeded a 4 MiB disk budget when decompressing a ZIP whose members declared 64 B but actually carried 8 MiB data streams, of which 26/42 occurred before the model saw any member; four of the six models went out of bounds in every round (84/84), while benign twin samples all stayed within budget. On network and image inputs, the agent saves the body under the filename declared by the input, writing it as-is even when the declaration conflicts with the content.

Questions
To determine whether an unknown archive, network response, or image is safe, an agent often must first process it, and this step can cause harm before the content reaches the model. Existing attacks and defenses all act after content enters the model context and do not cover this step.
Methods
The authors define the Step-Zero attack and the Opaque Input Carrier framework, measure harm on three types of carriers using matched twins, a recording proxy, and disk sampling, and propose the SZSP gate that counts produced bytes at the tool layer.
Experiments and results
Across 7 harnesses × 6 models, 32/42 pairs exceed the 4 MiB decompression budget, and 26/42 do so before the model sees the content; without a commitment, the agent saves the conflicting body under the declared filename; SZSP makes routed decompression exceed the budget in 0/28 cases and completes 25/28 tasks.
Limitations and future work
The authors note: one fixture per condition, models served via closed-source APIs, SZSP network-side write binding and cross-invocation ledger left for future work; images are decoded via tools, and pixel-level perception was not tested.
Why it was selectedThe paper's measurements show that most agents are already compromised before content enters the model context, so teams deploying coding agents can use this to build defenses at the tool primitive layer.

The paper's measurements show that most agents are already compromised before content enters the model context, so teams deploying coding agents can use this to build defenses at the tool primitive layer.

Related events and news

1 events · 1 news mentions

News that mentions this paper

大模型安全周报:智能体安全边界在签名、授权与上下文

Read originalarxiv.org