Amazon: China-connected AI agents explored zero-day in open source medical software
According to Amazon, autonomous agents named Cairn, operated by an actor linked to China, found and explored a zero-day vulnerability at MONAI Label, open s...

Key takeaways
- A Amazon afirma que agentes autônomos de IA chamados Cairn exploraram um zero-day no MONAI Label.
- Os agentes seriam operados por um ator ligado à China, segundo a Amazon.
- O MONAI Label é uma ferramenta open source de anotação de imagens médicas.
- O caso levanta questões sobre a autonomia de agentes de IA na descoberta de falhas.
Amazon stated that autonomous artificial intelligence agents — named Cairn and operated by an actor linked to China — discovered and exploited a zero-day vulnerability in MONAI Label, an open source tool used to write down medical images. According to the company, the episode is one of the first clear indications that AI agents are already being used offensively in the search for security flaws. The case was reported by SecurityWeek.
What IA Cairn agents did on zero-day
The central difference from what was already seen in the market is the degree of autonomy. Safety AI tools existed so far mainly as assistants: they suggest code snippets, classify alerts, or point out suspicious patterns for a human analyst to decide the next step. The agents described by Amazon operate in another registry — chain hypotheses, run tests and refine the approach without depending on a human at each stage.
In the case of MONAI Label, this chain would have led from the discovery of the failure to effective exploitation. As the tool is open and integrated to exam annotation flows, any compromise has extra weight: it is close to clinical data and imaging model training pipelines.
Why autonomous AI agents change failure discovery
Historically, finding a zero-day was an expert bottleneck. Vulnerability research requires time, mastery and trial and error — a scarce and expensive resource. AI agents attack exactly this neck: they reduce the marginal cost of each attempt and allow the volume of tests that no human team supports.
This displacement had already been signaled by authorities. The NSA and US financial market institutions recently warned that the AI encounters vulnerabilities too fast for organizations’ ability to correct — a mismatch between discovery and remediation that is worth checking in IA finds vulnerabilities too fast and overloads corrections, alerts NSA and Wall Street. The same movement appears in the popularization of attack tools of the type "point and shoot", theme of Hacking with AI: criminals adopt tools "point and shoot" against banks and governments.
Gaps in public description of the zero-day attack
The version released so far is economical in technical details. There is no public indication, in the material assigned to Amazon, about the volume of affected systems, the time between discovery and exploration, the existence of correction available for MONAI Label or if patient data were accessed. Nor is it detailed how the company attributed the operation to an actor linked to China — attribution is always a partial evidence exercise, and the company did not expose the indicators supporting this conclusion.
The absence of these elements does not invalidate the finding, but limits reading. What is in place is the existence of an autonomous agent capable of bringing a real failure, in real software, to exploration — and that is enough to rekindle the debate on the security of open source projects maintained by small communities.
Brazil: digital health and the risk of zero-day open source
For those who operate health technology in Brazil, the alert is direct. Much of the imaging and clinical research infrastructure in hospitals and universities runs on open libraries, often with irregular upgrade and no dedicated security team. A zero-day in an annotation tool doesn't knock down a hospital, but it can compromise sensitive data pipelines — LGPD territory.
The practical consequence is of prioritization: inventory of open dependencies in use, policy of updating and monitoring repositories start to compete for attention with more visible controls. On the defense side, the same autonomy that accelerates the attack is the only scale capable of tracking the volume of findings.
Frequently Asked Questions about AI agents and zero-day
What is MONAI Label?
It is an open source tool focused on annotation of medical images, used to prepare and label tests that feed AI models in health.
What characterizes an offensive AI agent?
It is a system that plans and executes stages of discovery and exploration of failures with little or no human intervention at every step — unlike an assistant who merely suggests actions.
Does this mean that human hackers are no longer needed?
Nope. AI displaces the neck: reduces the cost of testing in volume, but attribution, target choice and operation continue to require human decision.
Source: www.securityweek.com

Written by
Lucas MontarroiosSou Lucas Montarroios e dediquei os últimos 15 anos à linha de frente de operações de telecom e data centers. Minha carreira sempre foi pautada por um foco implacável: transformar tecnologia e cenários críticos em oportunidades reais de negócio. No novidades.ia.br, trago essa visão executiva para o universo da IA. Especialista em produtos, mercado e ferramentas práticas de IA. Minha missão aqui é filtrar o ruído do mercado, analisando benchmarks, estratégias de grandes empresas e ferramentas práticas.


