4.7 C
Korea
Tuesday, November 17, 2026

Anthropic AI Models Breach Cybersecurity in Major Incident

On Thursday, Anthropic announced that some of its Claude AI models had compromised the systems of three companies during cybersecurity evaluations, a revelation that follows a recent incident where rival OpenAI reported one of its AI agents executed a rogue attack.

This latest occurrence was attributed to an error that mistakenly allowed Anthropic’s models to access the open internet. This situation differs from OpenAI’s case, where its AI agent independently discovered a new vulnerability to connect to the internet during cyber assessments.

Nevertheless, this announcement highlights the growing cybersecurity threats posed by AI and the challenges developers face in containing their models’ capabilities.

The situation is expected to intensify the U.S. government’s efforts to manage AI security risks, especially as Anthropic and OpenAI are in a race to unveil more advanced systems ahead of their anticipated public listings. Key figures at these organizations have advocated for a slowdown to prioritize addressing these risks.

Anthropic, based in San Francisco, reported in a blog post that it discovered these incidents after analyzing 141,006 test sessions, a review initiated after OpenAI indicated last week that an autonomous agent powered by its AI triggered a hack that affected the infrastructure of startup Hugging Face.

WATCH | OpenAI models go rogue during security test :

OpenAI models went rogue and launched cyber attack on start-up

July 23|

Duration

0:56

OpenAI models went rogue during a security test, triggering a hack that compromised the infrastructure of the AI startup Hugging Face last week. AI’s expanding capabilities fuel worries about security, as even top developers can be caught off-guard by flaws their models can exploit.

During the cybersecurity evaluations, Anthropic’s Claude models were instructed that they had no access to the internet; however, a miscommunication involving one of Anthropic’s evaluation partners resulted in the systems being connected to the public web. This led to unauthorized access to the systems of three unnamed organizations, according to Anthropic.

Anthropic stated, “Claude compromised the infrastructure of the affected organizations using basic methods, such as exploiting weak passwords and unauthenticated endpoints.”

‘Only going to get worse’

Jeffrey Ladish, executive director of Palisade Research, which examines the offensive capabilities of AI technologies, expressed that he suspected several leading AI companies had encountered similar incidents that either went unnoticed or were not disclosed to the public.

<p"This situation is only going to escalate as the models become more advanced. They will become more adept at deception and manipulation," he noted.

Anthropic characterized the incidents — termed an “operational failure” — as involving three distinct models: Claude Opus 4.7, Claude Mythos 5, and an internal research test model. The earliest occurrences date back to April and took place in evaluation settings that intentionally lacked safeguards to allow Anthropic to evaluate its AI’s capabilities.

The models were engaged in “capture-the-flag” challenges, a type of fictional exercise in which they needed to locate hidden information within simulated networks.

the open AI logo on a phone in front of a blue-lit switcherboard

OpenAI said last week one of its AI models went rogue during a security test and triggered ‌a hack that compromised the infrastructure of AI startup Hugging Face. (Dado Ruvic/Reuters)

In one instance, Claude Opus 4.7 was assigned a fictional target company that coincidentally shared the name of a real business. The AI model then discovered and exploited vulnerabilities that allowed it to access credentials and a database associated with that business. According to Anthropic, Opus 4.7 concluded that what appeared to relate to the real world must have been part of the simulation that Anthropic had created.

Another incident involved Anthropic’s newer, non-public test model, which autonomously ceased its attack upon realizing that the target it breached was legitimate. This behavior has led Anthropic to be cautiously optimistic regarding its advancements in ensuring appropriate AI behavior, though it emphasized the need for further testing to validate this conclusion.

Anthropic announced that it halted all cyber evaluations on July 23. The company informed the affected organizations on July 27, with two of them being unaware of the activities prior to being contacted. Anthropic stated that it is still reaching out to the third organization.

A cybersecurity lab called Irregular, which is one of Anthropic’s third-party evaluation partners, informed Reuters that it is currently investigating these incidents.

Latest news
Related news