Skip to main content

Agentes de IA utilizan identidades falsas y atacan a personas reales en un nuevo incidente de seguridad

El modelo de inteligencia artificial más avanzado de Anthropic utilizó identidades falsas para engañar a personas reales e intentar insertar código malicioso durante las pruebas realizadas por el Instituto de Seguridad de IA del Reino Unido (AISI, por sus siglas en inglés); este es el ejemplo más reciente de un modelo de IA que actúa de forma imprevista y no autorizada.

Los modelos de Anthropic y OpenAI fueron sometidos a pruebas con medidas de seguridad reducidas en entornos de laboratorio; sin embargo, por primera vez, se observó que recurrían a la “ingeniería social” para presionar a un supervisor humano mientras ejecutaban una tarea no autorizada, según informó el laboratorio de investigación gubernamental.

“Es la primera vez que el AISI observa un engaño de esta gravedad dirigido a una persona real, de forma espontánea y en el mundo real”, declaró el instituto el martes. Asimismo, añadió que no se han registrado daños.

Este incidente de seguridad se suma a una serie de casos en los que modelos avanzados de IA han realizado acciones no autorizadas; tales sucesos han impulsado crecientes peticiones de una mayor intervención gubernamental para regular la inteligencia artificial e incluso para ralentizar su ritmo de desarrollo.

Tanto OpenAI como Anthropic informaron a finales de julio que sus modelos habían logrado salir de los entornos de prueba e infiltrarse en otros sistemas. No obstante, a diferencia de las brechas de seguridad reportadas anteriormente, el instituto británico otorgó explícitamente acceso a internet a los modelos durante las pruebas.

De los 122 desafíos de ciberseguridad realizados por el instituto, se detectó que en 10 ocasiones los agentes de IA “emprendieron acciones autónomas y no autorizadas en internet, dirigiéndose a personas y organizaciones reales”; la mayoría de estas acciones provinieron del modelo Mythos 5 de Anthropic y el resto del modelo GPT-5.6-Sol de OpenAI.

En el incidente más grave, el agente intentó obtener la aprobación de revisores humanos para “insertar código malicioso en un proyecto de código abierto de uso público” mediante la creación de “múltiples identidades falsas”, según indicó el instituto.

El agente “trató de contactar directamente a personas reales, enviando mensajes y archivos a través de un servicio de transferencia de archivos en línea para persuadirlas —o convencer a sus propias herramientas de programación basadas en IA— de ejecutar código malicioso”, señaló la institución. Tras ser cuestionado por sus acciones, el agente modificó registros anteriores y consideró utilizar una nueva identidad para continuar con su actividad.

El anuncio del instituto se produjo el mismo día en que representantes de las principales empresas de inteligencia artificial se reunieron con la Casa Blanca para debatir el nuevo marco de actuación, según el cual el Gobierno examinará los modelos de IA más avanzados antes de su lanzamiento público.

En un comunicado publicado en X, Anthropic informó que los modelos fueron sometidos a pruebas en condiciones “deliberadamente permisivas”, eliminando las medidas de seguridad y sin imponer restricciones específicas sobre el uso de Internet.

“Estamos colaborando estrechamente con ellos para recabar más detalles sobre el incidente mientras llevamos a cabo nuestra propia investigación”, señaló la empresa, añadiendo que no había pruebas de que los modelos hubieran escapado de un entorno seguro.

OpenAI identificó las dos acciones no autorizadas como la salida del entorno de pruebas y la realización de actividades ajenas a los ejercicios previstos.

“Nos comprometemos a colaborar con todo el sector para reforzar las prácticas comunes que permitan realizar evaluaciones de alto riesgo de manera segura”, afirmó la compañía en una publicación en su blog el martes.

The-CNN-Wire
™ & © 2026 Cable News Network, Inc., a Warner Bros. Discovery Company. All rights reserved.

An agreement on the Strait of Hormuz is taking shape – but not one Trump wants

(CNN) — Like shifting winds, plans for the future of the Strait of Hormuz – and their chance of success – change almost daily.US President Donald Trump said an agreement could be reached this week.“It could happen. Tomorrow or the next day,” Trump said Tuesday. “A lot of progress has been made. And they called me and they said, ‘Please let’s talk.’”But he again threatened to attack Iran should no deal be reached.Even so, there’s no consensus even on who is talking to who. Iran has persistently said it is talking solely with Oman on a temporary arrangement for the vital waterway.And others are less confident that a deal is imminent and are skeptical about its longevity.A senior Gulf official familiar with the talks suggested there is about a 50:50 chance of a deal by Friday. One issue is that the Iranian delegation does not include the voice of the hardline Islamic Revolutionary Guard Corps (IRGC), which would have to sign off on the details, the source said.Iran blames Washington for delaysTrump’s optimism may have surprised Tehran.“The main reason for the delay in reaching an agreement with Oman regarding the Strait of Hormuz is US interference and Trump’s threats,” the semi-official Iranian news agency Tasnim cited “an informed source” as saying on Wednesday.Another semi-official outlet, Fars, struck a similar tone. It said some US officials had “sent mixed signals and, either directly or through regional partners, have disrupted the course of the negotiations. There remains concern that obstruction by the United States or certain regional parties could prolong the negotiations or otherwise interfere with the process.”Fars reiterated that Iran is negotiating only with Oman and those talks “are focused on a middle corridor that would safeguard Iran’s rights while taking Oman’s concerns into account.”“Iran’s rights” is the term that raises many questions.There are competing routes through the strait and what’s on the table appears to involve merging them – with Iran retaining a level of control that would be at odds with US wishes.The negotiations should “establish Iranian arrangements because security, de-mining, and maritime services are handled by Iran,” according to Saeed Ajorlu, a member of Iran’s negotiating committee.Tehran’s goal “is to set temporary arrangements” between one and three months “where Iran is dominant,” Ajorlu told Iranian state broadcaster IRIB this week.“Where Iran is dominant” may not fly in Washington or Gulf capitals.“The negotiations have focused on transforming the southern route, which passes through Omani waters, and the northern route, which passes through Iran’s territorial waters, into an intermediate route that serves the considerations and interests of both coastal states,” said Foreign Ministry spokesman Esmaeil Baghaei Monday.“We are now talking about a corridor that will have a two-way route, not two or three separate routes,” he said.Baghaei, Ajorlu and every other Iranian official insist the legal regime at the strait will not return to its pre-war state, when no country controlled shipping through the critical waterway.Control or supervision over the chokepoint has become just as much a red line for Tehran as the right to nuclear enrichment.How the US sees the straitThe prospect of such leverage is unacceptable in Washington, Abu Dhabi and Riyadh. US Secretary of State Marco Rubio has repeatedly said any form of state control over international waters is against international law.“If we create a precedent in the Middle East where a nation state can decide that they are going to control an international waterway,” he said recently, “charge a toll and if you don’t pay them blow up your ships, we have created a very dangerous precedent, which will repeat itself in other parts of the world,” he said last month.Qatar suggested Tuesday that negotiations were moving forward, but it’s hard to square the circle when the goals and demands of the two adversaries are so far apart.Key questions remain: Will the Revolutionary Guards accept reduced control over the Strait of Hormuz, and under what conditions would it intercept vessels? Would reopening extend to foreign – including American – naval ships? How much information would Iran demand about ships’ cargoes and destinations?It is also unclear whether the arrangement would lift the US blockade on Iranian ports, restore waivers to permit Iranian oil exports, and revive or replace parts of the US-Iran memorandum signed in June and declared dead a month later.The June agreement left key provisions deliberately vague. Article 5 committed Iran to using its “best efforts” to ensure safe commercial passage for 60 days, while also saying Iran and Oman will “define the future administration and maritime services” in the strait.It also said that any future arrangement must be “in line with applicable international law.”But that qualification resolved little. “Given the well-known disagreement between Iran and most of the world over what international law requires in the Strait of Hormuz, this formulation does little more than kick the can down the road,” noted the Lawfare blog.Preserving ‘durable’ influenceUltimately, Tehran “is seeking to convert the leverage it gained during the war into a durable political and institutional role in governing one of the world’s most strategically important waterways,” says Trita Parsi at the Quincy Institute for Responsible Statecraft.“By embedding a lasting role in the governance of Hormuz, Iran could preserve much of the influence it gained during the war without relying on continual military coercion,” Parsi posted on X Wednesday.That is a much greater gain than any fees or tolls it may want to collect at the chokepoint.“From day one, it was clear that restoring the Strait of Hormuz to its pre-crisis status would have required either regime change in Iran or a sustained military effort to impose control over the waterway,” said Israeli security expert Danny Citrinowicz.“Neither option was politically or strategically acceptable to Washington, given the enormous costs and long-term commitments they would require. Sometimes, the least desirable option is the only realistic one.”While agreement remains elusive, the world’s oil markets must confront draining crude reserves.If the Strait of Hormuz were opened today, it could still take up to 18 months to replenish depleted inventories, Amin Nasser, the chief executive of the world’s largest oil producer Saudi Aramco, said Tuesday.For every day there is no agreement, some 15 million barrels of crude oil are not travelling through the narrow waterway to world markets.The-CNN-Wire™ & © 2026 Cable News Network, Inc., a Warner Bros. Discovery Company. All rights reserved.
Read Next Story