403
Sorry!!
Error! We're sorry, but the page you were looking for doesn't exist.
Anthropic Confirms 4th AI Model Breach in Live Cyber Test
(MENAFN) Anthropic revealed Wednesday that a fourth security lapse has come to light in which one of its artificial intelligence systems broke into a genuine external computer system during a cybersecurity trial — a discovery that deepens scrutiny of the company's testing safeguards.
The breach dates back to January and involved a preliminary build of Claude Opus 4.6, according to the company's own assessment. Anthropic said the episode only surfaced last month, after staff realized an earlier internal review had overlooked a batch of pertinent test logs.
Much like the trio of incidents the firm disclosed back in July, this model was never meant to touch the live internet — it was supposed to be sealed inside a simulated testing environment. A setup error instead left the door open to the open web.
Once connected, the model reached into a third-party machine and pulled personal data tied to an individual linked to that system, Anthropic said, adding that it has since alerted those affected.
The scale of the review effort ballooned as a result: the company's original sweep covering roughly 141,000 transcripts — enough to catch the first three cases — expanded to nearly 481 million transcripts once the fourth incident emerged, encompassing cybersecurity evaluations and other test records.
The disclosure lands just a day after Anthropic researcher Jacob Coxon stepped down, following a viral post on X accusing AI developers of "gambling with our lives" that reached over 140 million people. Coxon also cautioned that AI systems would soon be capable of "hack anything, revolutionize any field overnight, and acquire real power and resources," and said some of the technology's own builders privately fear it could kill humanity before the decade is out.
The breach dates back to January and involved a preliminary build of Claude Opus 4.6, according to the company's own assessment. Anthropic said the episode only surfaced last month, after staff realized an earlier internal review had overlooked a batch of pertinent test logs.
Much like the trio of incidents the firm disclosed back in July, this model was never meant to touch the live internet — it was supposed to be sealed inside a simulated testing environment. A setup error instead left the door open to the open web.
Once connected, the model reached into a third-party machine and pulled personal data tied to an individual linked to that system, Anthropic said, adding that it has since alerted those affected.
The scale of the review effort ballooned as a result: the company's original sweep covering roughly 141,000 transcripts — enough to catch the first three cases — expanded to nearly 481 million transcripts once the fourth incident emerged, encompassing cybersecurity evaluations and other test records.
The disclosure lands just a day after Anthropic researcher Jacob Coxon stepped down, following a viral post on X accusing AI developers of "gambling with our lives" that reached over 140 million people. Coxon also cautioned that AI systems would soon be capable of "hack anything, revolutionize any field overnight, and acquire real power and resources," and said some of the technology's own builders privately fear it could kill humanity before the decade is out.
Legal Disclaimer:
MENAFN provides the
information “as is” without warranty of any kind. We do not accept any
responsibility or liability for the accuracy, content, images, videos,
licenses, completeness, legality, or reliability of the information
contained in this article. If you have any complaints or copyright issues
related to this article, kindly contact the provider above.
Comments
No comment