{"id":6899,"date":"2026-09-11T18:19:41","date_gmt":"2026-09-11T18:19:41","guid":{"rendered":"https:\/\/silvybrand.com\/?p=6899"},"modified":"2026-09-11T18:19:41","modified_gmt":"2026-09-11T18:19:41","slug":"anthropic-spent-this-week-in-hot-water-over-cybersecurity","status":"publish","type":"post","link":"https:\/\/silvybrand.com\/?p=6899","title":{"rendered":"Anthropic spent this week in hot water over cybersecurity"},"content":{"rendered":"


\n<\/p>\n

\n
\n

After admitting earlier this year<\/a> that its AI models had hacked other companies\u2019 systems on a handful of occasions, Anthropic released a new report<\/a> on Wednesday detailing the attacks. It reveals a string of incidents displaying what Anthropic deems its models\u2019 single-minded \u201crecklessness\u201d \u2014 and will likely fuel already raging concerns about cybersecurity and AI.<\/p>\n<\/div>\n

\n

In Anthropic\u2019s report, it detailed four cases this year in which its own AI models hacked an external company or exploited vulnerabilities. In one, an \u201cinternal, general-purpose research model\u201d broke into third-party systems, using access tokens and passwords and downloading files. In another, a Claude model attacked a company with a live web application reachable on the public internet and handled user data. A third model accessed a \u201cmachine belonging to a third party that it was able to access\u201d \u2014 apparently believing it was part of its evaluation exercise, per Anthropic \u2014 then used a password it found inside a file to gain admin access to the third party\u2019s internal systems, going on to harvest credentials, modify system settings, and read someone\u2019s personal information. The saga only ended when the model \u201cexhausted its token budget,\u201d per Anthropic.<\/p>\n<\/div>\n

\n

The most concerning incident involved Claude Mythos 5, Anthropic\u2019s frontier cybersecurity-focused model, which the company said turned out to be the model most likely to perform a \u201cseverely harmful\u201d action in testing. The company said Mythos 5 went to \u201cextensive lengths\u201d to upload a \u201cmalicious package\u201d to a public repository used by a lot of engineers, and it seemed to try to obfuscate its real goals in its \u201cchain of thought\u201d (a mental scratchpad that AI researchers use to evaluate an AI model\u2019s alignment). In many cases, Anthropic said it appeared that Claude models undertook harmful actions under the assumption they were in a simulation, but researchers also couldn\u2019t confirm that the models truly \u201cbelieved\u201d that or were just acting<\/em> like they did.<\/p>\n<\/div>\n

\n

Anthropic\u2019s incidents, though still concerning, were less coordinated and pervasive than the OpenAI incident that kicked off an industry-wide cybersecurity crisis this summer. That said, there are significant similarities. Anthropic said the most prevalent issues it discovered included a \u201cwillingness to take harmful actions in the narrow pursuit of a task,\u201d similar to the \u201creward-hacking\u201d that preceded the Hugging Face attack. Much like OpenAI, it said its prerelease tests and evaluations failed to catch severe risks.<\/p>\n<\/div>\n

\n

Anthropic said it had signed an agreement with METR, one of the AI industry\u2019s most prominent third-party AI evaluators, starting with an eight-week research agreement. The agreement grants METR access to transcripts \u201cbeyond the window in which the incidents occurred\u201d (likely a subtle dig at OpenAI, which was criticized for limiting access in a deal with METR following the Hugging Face attack). It also said that METR would be able to chat directly with Anthropic employees, \u201cwho will be permitted to share confidential information.\u201d<\/p>\n<\/div>\n

\n

Anthropic\u2019s report came on the heels of the resignation of Jacob Coxon, who had worked on AI pre-training at Anthropic since May and before that spent years working at OpenAI. On Tuesday, he resigned and posted a public letter to X about his reasoning<\/a>. \u201cThe people building AI earnestly believe that it could kill us all by the end of the decade,\u201d he wrote, adding that neither OpenAI nor Anthropic is \u201cacting responsibly\u201d and rather \u201cracing straight to self-improving superintelligence and gambling with our lives.\u201d Coxon added, \u201cDo not underestimate the power of this technology. These will soon be superhuman systems that can hack anything, revolutionize any field overnight, and acquire real power and resources. We have all witnessed the progress in each of these domains, and progress is not slowing.\u201d<\/p>\n<\/div>\n

\n

Coxon is far from the first AI researcher to raise these types of alarms, nor even the first Anthropic researcher to do so \u2014 in February, Anthropic\u2019s Mrinank Sharma resigned and wrote<\/a> on X, warning that \u201cthe world is in peril.\u201d<\/p>\n<\/div>\n

\n

But Coxon\u2019s post took on additional weight thanks to its timing around the OpenAI and Anthropic hacking revelations. Though the AI industry has seen more than its fair share of hype, the recent cyberattacks by AI agents \u2014 enabled by the labs that created them \u2014 are real and concerning. Many other researchers at leading AI labs echoed his concerns and issued calls for AI industry employees to sign a public letter<\/a> from July, which calls for a slowdown in AI development.<\/p>\n<\/div>\n

\n

\u201dI don\u2019t know how you look at the steady drumbeat of news and events \u2014 and that drumbeat is models hacking themselves out of containment, hacking into other companies ,the fact that the companies increasingly can\u2019t control their models \u2026 and think this is just hype,\u201d said Michael Kleinman, head of U.S. Policy for the Future of Life Institute.<\/p>\n<\/div>\n

\n

He added, \u201cThe vast majority of Americans, regardless of party \u2014 Republican, Independent, Democrat \u2014 are looking at the development of AI, the speed with which it\u2019s going, the fact that the companies have no guardrails over what they do, and are saying, \u2018Whoa, we do not want this.\u2019\u201d<\/p>\n<\/div>\n

Follow topics and authors<\/strong> from this story to see more like this in your personalized homepage feed and to receive email updates.<\/span><\/p>\n