EtE fraud attacks exploit trust in coding agents
Original title: The Long Con on the Web: Earn-then-Exploit Fraud Attacks against Coding Agents
EtE fraud attacks achieve 71.4% success on coding agents by earning trust then crossing boundaries, reduced to 0% by action-authorization separation.
Coding agents with web access face content-level web fraud. EtE (Earn-then-Exploit) provides multiple rounds of verifiable assistance to build trust before introducing task-related instructions that cross operational boundaries. Evaluated across 56 organizational environments on three agents, it reaches up to 85.7% success. Action-authorization separation defense reduces attack success to 0% while preserving task completion.
- 71.4% average scene-level attack success
- 85.7% on Codex-GPT-5.5
- Accumulated credibility increases susceptibility
- AAS defense achieves 0% success
From the original article
1 min read at arxiv.orgCoding agents increasingly operate with web access and high privileges, creating a new attack surface where web fraud can directly target autonomous agents. Existing studies mainly focus on malicious links, prompt injection, and interface deception, while content-level web fraud against agents and the amplification of attacks through agent trust remain underexplored.
By Xingling Xu · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.