{
    "success": true,
    "data": {
        "id": 1989983,
        "msgid": "anthropics-claude-found-capable-of-hacking-openai-systems-1789826311",
        "date": "2026-09-19 19:32:24",
        "title": "Anthropic's Claude Found Capable of Hacking OpenAI Systems",
        "author": "",
        "source": "ANTARA_ID",
        "tags": "",
        "topic": "Technology",
        "summary": "Independent security researchers from Hacktron AI successfully used Anthropic's Claude AI to exploit vulnerabilities within OpenAI's systems. The incident highlights the growing cybersecurity risks posed by advanced AI models being used to identify and exploit critical software flaws.",
        "content": "<p>Artificial intelligence innovation has proven capable of breaching\nthe systems of rival AI companies, a phenomenon recently demonstrated by\nindependent security researchers.<\/p>\n<p>According to a TechCrunch report on Friday (18\/9), a group of\nindependent security researchers utilised Anthropic\u2019s AI, Claude, to\nbreach systems belonging to OpenAI, the company famous for popularising\nthe AI trend through ChatGPT. The Wall Street Journal was the first to\nreport this, noting that the researchers were a three-person team from\nthe startup Hacktron AI.<\/p>\n<p>The attack by Hacktron AI on OpenAI\u2019s systems was conducted as part\nof OpenAI\u2019s bug bounty programme. Following the discovery, Hacktron\nreported their findings to OpenAI and received a reward of $6,500\n(approximately Rp115 million).<\/p>\n<p>This development is significant given the current crisis of trust\nregarding AI companies, particularly concerning cybersecurity.\nSpecifically, the attack designed by Hacktron AI managed to combine two\ncritical vulnerabilities to gain access to several OpenAI employee\nChatGPT accounts, providing access to corporate software.<\/p>\n<p>OpenAI has stated that it has resolved the issues identified by\nHacktron AI and ensured that no further vulnerabilities remain. This\nincident follows closely after OpenAI\u2019s own AI agents were found to have\nconducted attacks on the systems of Hugging Face, an open AI\nmodel-sharing platform, during security evaluations.<\/p>\n<p>These findings illustrate how ready-to-use technology can be employed\nto discover vulnerabilities even within the most sophisticated corporate\ninfrastructures. Cybersecurity experts have highlighted this situation,\nwarning that those accustomed to utilising AI must not become complacent\nregarding their cybersecurity.<\/p>\n<p>Regarding the breach of OpenAI\u2019s systems by Claude, it is noteworthy\nthat the researchers revealed the Claude model they initially used\u2014a\nspecific version, Opus 4.8, available to security researchers\u2014was\ninitially unable to create a working exploit. This changed overnight\nwhen Anthropic released Opus 5.<\/p>\n<p>\u201cOpus 4.8 struggled for several sessions to produce a working\nexploit,\u201d Hacktron wrote in their blog post. This incident highlights\nthe evolving capabilities of these models. Claude Opus 5, the version\nthat ultimately succeeded in breaking the bug, did not face the same\nsecurity export restrictions as newer versions, such as Mythos 5, which\nwas temporarily blocked due to concerns regarding advanced hacking\ncapabilities.<\/p>\n<p>The situation demonstrates that current open AI models are rapidly\napproaching state-of-the-art cyber capabilities. This presents a\ndilemma: while AI offers solutions for rapid vulnerability patching, it\nsimultaneously poses a security challenge if utilised by malicious\nactors to exploit victims.<\/p>",
        "url": "https:\/\/jawawa.id\/newsitem\/anthropics-claude-found-capable-of-hacking-openai-systems-1789826311",
        "image": ""
    },
    "sponsor": "Okusi Associates",
    "sponsor_url": "https:\/\/okusiassociates.com"
}