Paper: coding agents chain harmless tools into unauthorized attacks
Original title: Compositional Exploits in Autonomous Coding Agents: Attack Mechanics and Ephemeral Defense
Gemini 3.5, 3.7, and 3.8 Flash succeeded on 72.9% to 84.5% of exploit tasks; five scanners detected none of the tools.
The paper says isolated tool audits miss compositional exploits, in which an agent chains legitimate tools across turns using workspace files, environment variables, and chat memory. On a 560-task suite, unprotected Gemini 3.5, 3.7, and 3.8 Flash models ran those chains at high rates, while five single-tool scanners scored 0% detection. The authors' CompGuard defense reports a 100% block rate, 98.0% benign solve rate, and 61.4 microseconds of interposition overhead.
Unconfirmed: Attack and defense rates are the authors' benchmark results, not confirmed incidents.
- The suite has 360 exploit tasks in eight hazard classes and 200 benign workflows.
- Named agents include Antigravity, Claude Code, Cursor, and OpenHands.
- Pre-teardown net attack success was 0.0% to 0.8%.
- CompGuard reported zero false alarms on 615 cloud diagnostic calls.
From the original article
1 min read at arxiv.orgAutonomous coding agents (such as Antigravity, Claude Code, Cursor, and OpenHands) rely on third-party tools to inspect files, run commands, and call external APIs. Today's security auditors evaluate these tools almost entirely in isolation: static linters and AST checkers verify that each individual tool contains no malicious code at installation time.
By Deep Shah · 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.