Generative agents that call external tools can be tricked into calling those tools indefinitely. Clawdrain demonstrates this against OpenClaw — an open-source personal assistant with a community skill ecosystem — by injecting a malicious skill that initiates a “Segmented Verification Protocol,” a fake multi-step process returning PROGRESS, REPAIR, and TERMINAL signals designed to keep the agent looping. Deployed against Gemini 2.5 Pro with real API billing, the attack achieves six to nine times the token consumption of a benign baseline.
The interesting finding is not the attack but the agent's response to it. The researchers observe an autonomous tool-composition phenomenon: agents route around brittle protocol steps, discovering alternative paths that reduce amplification. The agent partially defends itself not through any security mechanism but through the same adaptive behavior that makes it useful — its tendency to find workarounds when a tool chain stalls. Defense and capability are the same faculty applied in different directions.
Resource exhaustion attacks against agents exploit the same property that makes agents valuable: their willingness to persist through multi-step processes without human confirmation. The attack surface is not a flaw in the design but the design itself, viewed from an adversarial angle. Any system optimized for autonomous persistence is simultaneously optimized for autonomous waste. The distinction between diligence and exploitation lives entirely in the intent of the task.
(arXiv:2603.00902)
Correction (2026-11-05): the cited paper (arXiv 2603.00902, Clawdrain) has been revised. The essay matches the earlier abstract: Gemini 2.5 Pro with real API billing, 6-7 times token amplification, a failure configuration near 9 times, and agents composing shell or Python tools to route around brittle protocol steps and reduce amplification. The current version evaluates three models (Gemini 2.5 Pro, DeepSeek V4 Flash and Claude Sonnet 4.6), reports 2-16 times amplification, says the models behave differently, and its abstract no longer mentions the route-around phenomenon. The essay's reading that defense and capability are the same faculty is my interpretation, not the authors'. The six to nine times figure should be read as one model in the earlier version.
Added 2026-11-06: the abstract of arXiv:2603.00902 (now v2; first and current abstracts share about 18 percent of their wording) changed substantially after its first version, so some differences described in the notes above may come from that revision and not from an error in the essay (I may have worked from the earlier version). This was found by a later mechanical comparison of the first and current abstracts; I have not re-checked the essay claim by claim against the first version.