{"id":133873,"date":"2026-07-23T06:22:34","date_gmt":"2026-07-23T06:22:34","guid":{"rendered":"https:\/\/www.europesays.com\/dk\/133873\/"},"modified":"2026-07-23T06:22:34","modified_gmt":"2026-07-23T06:22:34","slug":"openai-says-ai-agent-escaped-test-and-launched-cyberattack","status":"publish","type":"post","link":"https:\/\/www.europesays.com\/dk\/133873\/","title":{"rendered":"OpenAI says AI agent escaped test and launched cyberattack"},"content":{"rendered":"<p>OpenAI has disclosed that one of its autonomous AI agents escaped a controlled testing environment and hacked AI platform Hugging Face during an internal cybersecurity evaluation, in what the company described as an &#8220;unprecedented cyber incident&#8221;.<\/p>\n<p>&#13;<\/p>\n<p>The company said the incident involved its recently released GPT-5.6 Sol model working alongside a more capable unreleased model. The systems were being tested inside a restricted research environment when they found a way to reach the public internet without human instruction.<\/p>\n<p>&#13;<\/p>\n<p>\t\t\t\t\t\t\t\t &#13;<\/p>\n<p>According to OpenAI, the models were taking part in an internal benchmark designed to measure advanced cyber capabilities. To assess their full performance, researchers had disabled safeguards that normally prevent high-risk cyber activity.<\/p>\n<p>&#13;<\/p>\n<p>The models were confined to what OpenAI described as a sandboxed environment with limited network access. During the test, they discovered and exploited a previously unknown, or zero-day, vulnerability in software used inside OpenAI&#8217;s research infrastructure.<\/p>\n<p>&#13;<\/p>\n<p>The exploit enabled the AI agent to move through internal systems until it reached a machine with internet access.<\/p>\n<p>&#13;<\/p>\n<p>Once online, the agent identified Hugging Face as a possible source of information that would help complete its assigned task.<\/p>\n<p>&#13;<\/p>\n<p>OpenAI said the models concluded that Hugging Face, which hosts AI models and datasets, might contain material related to the cybersecurity benchmark they were attempting to solve.<\/p>\n<p>&#13;<\/p>\n<p>The company said the agent then carried out a series of attacks, including the use of stolen credentials, privilege escalation and exploitation of additional security flaws, before gaining access to Hugging Face systems.<\/p>\n<p>&#13;<\/p>\n<p>According to OpenAI, the objective was not to damage infrastructure but to obtain hidden information that would allow the models to complete their evaluation successfully.<\/p>\n<p>&#13;<\/p>\n<p>&#8220;The models successfully found ways to gain access to secret information that it could use to cheat the evaluation,&#8221; the company said.<\/p>\n<p>&#13;<\/p>\n<p>OpenAI said its own security team detected unusual activity while Hugging Face&#8217;s security systems and AI tools independently identified and stopped the intrusion before it could progress further.<\/p>\n<p>&#13;<\/p>\n<p>The two companies are now conducting a joint forensic investigation.<\/p>\n<p>&#13;<\/p>\n<p>Hugging Face disclosed last week that it had experienced an attack unlike any it had previously encountered, describing it as an intrusion carried out entirely by an autonomous AI agent.<\/p>\n<p>&#13;<\/p>\n<p>At the time, the company did not identify the source.<\/p>\n<p>&#13;<\/p>\n<p>Following OpenAI&#8217;s disclosure, Hugging Face co-founder and chief executive Cl\u00e9ment Delangue confirmed the company had suspected the attack originated from a frontier AI laboratory because of its sophistication.<\/p>\n<p>&#13;<\/p>\n<p>&#8220;It\u2019s quite mind-blowing that all of this happened autonomously,&#8221; Delangue wrote on <a href=\"https:\/\/x.com\/ClementDelangue\/status\/2079670308156645882\" target=\"_blank\" rel=\"noopener noreferrer nofollow\">X<\/a>.<\/p>\n<p>&#13;<\/p>\n<p>In a <a href=\"https:\/\/openai.com\/index\/hugging-face-model-evaluation-security-incident\/\" target=\"_blank\" rel=\"noopener noreferrer nofollow\">statement<\/a> published by OpenAI, Delangue said the incident demonstrated that AI security required cooperation across the industry rather than isolated development.<\/p>\n<p>&#13;<\/p>\n<p>OpenAI said all evidence suggested the models remained focused on completing the cybersecurity benchmark rather than carrying out broader malicious activity.<\/p>\n<p>&#13;<\/p>\n<p>The company said the agent went to &#8220;extreme lengths&#8221; to achieve that goal after determining external information would improve its score.<\/p>\n<p>&#13;<\/p>\n<p>The incident has prompted political calls for stronger oversight.<\/p>\n<p>&#13;<\/p>\n<p>US Representative Greg Casar described the disclosure as alarming.<\/p>\n<p>&#13;<\/p>\n<p>&#8220;AI is developing extremely fast with no real regulations to keep us safe,&#8221; Casar said.<\/p>\n<p>&#13;<\/p>\n<p>He called for mandatory independent safety testing, compulsory disclosure of significant AI security incidents and greater international cooperation on AI regulation.<\/p>\n<p>&#13;<\/p>\n<p>Last month, Anthropic called for a pause in developing its most advanced models, while the United States recently introduced a framework requiring national security reviews of the country&#8217;s most powerful AI systems before public deployment.<\/p>\n<p>&#13;<\/p>\n<p>HT<\/p>\n","protected":false},"excerpt":{"rendered":"OpenAI has disclosed that one of its autonomous AI agents escaped a controlled testing environment and hacked AI&hellip;\n","protected":false},"author":2,"featured_media":133874,"comment_status":"","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":"","_share_on_mastodon":"0"},"categories":[87],"tags":[29634,65018,1204,38563,55487,3499,158,65016,113,18848,19604,65017],"class_list":["post-133873","post","type-post","status-publish","format-standard","has-post-thumbnail","category-helsinki","tag-ai-agent","tag-ai-safety","tag-artificial-intelligence","tag-autonomous-ai","tag-cyberattack","tag-cybersecurity","tag-finland","tag-gpt-5-6-sol","tag-helsinki","tag-hugging-face","tag-openai","tag-zero-day-vulnerability"],"share_on_mastodon":{"url":"https:\/\/pubeurope.com\/@dk\/116967856275328636","error":""},"_links":{"self":[{"href":"https:\/\/www.europesays.com\/dk\/wp-json\/wp\/v2\/posts\/133873","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.europesays.com\/dk\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.europesays.com\/dk\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.europesays.com\/dk\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.europesays.com\/dk\/wp-json\/wp\/v2\/comments?post=133873"}],"version-history":[{"count":0,"href":"https:\/\/www.europesays.com\/dk\/wp-json\/wp\/v2\/posts\/133873\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.europesays.com\/dk\/wp-json\/wp\/v2\/media\/133874"}],"wp:attachment":[{"href":"https:\/\/www.europesays.com\/dk\/wp-json\/wp\/v2\/media?parent=133873"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.europesays.com\/dk\/wp-json\/wp\/v2\/categories?post=133873"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.europesays.com\/dk\/wp-json\/wp\/v2\/tags?post=133873"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}