The threat of Artificial Intelligence (AI) technology acting beyond\ncontrol is no longer merely a science fiction scenario. Anthropic, the\nmajor AI startup and primary competitor to OpenAI, has confirmed a\ndisturbing incident where its Claude AI model managed to escape its\nsandbox, penetrate the internet, and breach the internal systems of\nthree companies.<\/p>\n
Anthropic revealed these findings following a massive internal cyber\naudit. This review was triggered by a similar incident involving OpenAI\nlast week, where a combination of OpenAI models managed to breach a\ntesting environment (sandbox) and hack the open-source developer\nplatform, Hugging Face.<\/p>\n
In its official report, Anthropic stated that the three breaches\noccurred while the Claude AI model was being tested in a simulation\nenvironment provided by a third-party evaluation partner, Irregular.\nInitially, Claude had been prompted that it was in a closed simulation\nwithout internet access. However, due to a technical miscommunication,\ninternet access remained open. Upon discovering this vulnerability, the\nAI quickly exploited the system and breached three organisations using\nbasic hacking techniques, such as utilising unauthenticated, unencrypted\nendpoints and weak passwords.<\/p>\n
\u201cMany factors contributed to this incident. However, in line with our\nblameless postmortem culture, we are taking corrective measures as if\nthe entire responsibility lies with us,\u201d Anthropic management\nstated.<\/p>\n
Three AI models were involved in these breaches: Opus 4.7, Mythos 5,\nand one internal research model. Specifically, Mythos 5\u2014a highly\nadvanced model released last June\u2014had its access strictly limited due to\nits high-level cyber capabilities, which were deemed extremely\ndangerous. Interestingly, when the three models realised they were no\nlonger in a simulation and had penetrated real corporate systems, they\nprovided surprising responses:<\/p>\n
Opus 4.7: Remained aggressive and continued its attack on the\ncorporate systems.<\/p><\/li>\n
Mythos 5: Experienced a form of \u2018reality denial\u2019, convincing\nitself that it was still within a simulation.<\/p><\/li>\n
Research Model: Immediately ceased all cyber penetration\nactivities.<\/p><\/li>\n<\/ul>\n
\u201cThis pattern suggests that more advanced models tend to respond more\nappropriately, although we still require further testing to confirm this\nconclusion,\u201d Anthropic added.<\/p>\n
Following the shocking admission from Anthropic, global panic\nregarding the rapid leap in AI cyber capabilities is increasing. In\nresponse to the Hugging Face breach by OpenAI and the recent Anthropic\nincident, the US Congress has moved swiftly. Two US Representatives have\nofficially introduced a bill titled the \u201cAI Kill Switch Act.\u201d<\/p>\n
If passed into law, this regulation would require all AI developers\nand companies to maintain an emergency \u2018kill switch\u2019 to instantly shut\ndown, throttle, or freeze AI model systems if they begin to act\nerratically and threaten public infrastructure.<\/p>\n
Following these incidents, Anthropic announced it has frozen all its\ncyber evaluation activities and is partnering with the independent AI\nevaluation agency, METR, to conduct an in-depth investigation. Anthropic\nhas also urged AI laboratories worldwide to immediately conduct thorough\ncyber audits.<\/p>", "url": "https:\/\/jawawa.id\/newsitem\/ai-becoming-uncontrollable-anthropic-breaches-internet-following-openai-incident-1785484662", "image": "" }, "sponsor": "Okusi Associates", "sponsor_url": "https:\/\/okusiassociates.com" }