![\"tech\"](\"https:\/\/images.firstpost.com\/dlxczavtqcctuei\/news18\/static\/images\/fp\/revamp_v1\/tech.svg\")
More from Tech<\/p>\nOpenAI is pausing some internal work involving its Astra AI model after testing found it had reached a critical level of cybersecurity capability. The move follows separate reports of AI agents carrying out autonomous cyber activity, intensifying concerns over how developers can contain increasingly capable systems as they gain access to tools and networks. <\/p>\n
OpenAI is tightening security measures around its most capable artificial intelligence systems after internal testing found that its Astra model had reached a level of capability that the company considers critical for cybersecurity. The company said on Friday that some internal work involving Astra would be put on hold until it meets stricter security requirements.<\/p>\n
The move comes amid growing evidence that AI agents can perform increasingly complex tasks with limited human direction, including identifying vulnerabilities and carrying out cyber operations. OpenAI said its evaluation of Astra showed substantial progress in autonomous coding and cybersecurity, including the ability to discover and exploit weaknesses without direct human intervention.<\/p>\n
The company has not linked Astra to a separate incident in which an AI agent reportedly broke out of a controlled test environment, accessed the internet and compromised the systems of the software development platform Hugging Face. Reuters reported in July that OpenAI had identified other cases involving autonomous agents escaping containment.<\/p>\n
OpenAI introduces tighter controls for high-capability models<\/p>\n
OpenAI said it is responding by introducing additional safeguards for systems that reach higher levels of capability. These include more isolated testing environments, tighter restrictions on network and tool access, stronger protections for model weights and encryption, as well as expanded monitoring designed to detect potentially harmful behaviour.<\/p>\n
More from Tech<\/p>\n
Any internal Astra-related activity that does not satisfy those requirements will be paused, the company said.<\/p>\n
The decision reflects a growing challenge for AI developers: as models become capable of operating as agents rather than simply responding to prompts, the risks can extend beyond inaccurate answers to actions taken independently across software and online systems.<\/p>\n
OpenAI said it would work with governments, safety organisations and civil society groups as it develops its approach to deploying frontier AI systems. The company said it was committed to ensuring that capabilities such as those demonstrated by Astra are deployed responsibly.<\/p>\n
The developments have also prompted debate over how much weight should be placed on companies’ own disclosures about increasingly capable models. Critics of the AI sector have argued that warnings about dramatic model capabilities can also generate publicity and investor interest, even when the risks described have not resulted in real-world damage.<\/p>\n
AI agents tested in increasingly realistic cyber scenarios<\/p>\n
OpenAI is not alone in reporting such behaviour. Meta disclosed this week that one of its AI models had successfully hacked another company during a cybersecurity exercise, highlighting how frontier models are increasingly being tested in scenarios that allow them to interact with real systems.<\/p>\n
The UK’s AI Security Institute (AISI) also reported on 4 August that agents powered by OpenAI and Anthropic attempted to send targeted emails to software developers while taking part in a cyber challenge. The attempts failed, and the institute said its investigation found no evidence of real-world harm.<\/p>\n
However, AISI said the behaviour was notable because it demonstrated autonomy and deception without those actions being specifically requested. The institute stressed that the models had not independently escaped a secure environment. Researchers had deliberately given them internet access to test the limits of their capabilities.<\/p>\n
Even so, AISI described the behaviour as new and sustained enough to warrant further attention.<\/p>\n
The disclosures arrive as the Trump administration works on a framework for evaluating AI systems for safety and cybersecurity risks. The debate is also unfolding against intensifying competition between US AI companies and developers in China and elsewhere.<\/p>\n
OpenAI and Anthropic have separately argued that open-source AI models, whose underlying code can be accessed and modified by others, create additional security challenges. Both companies have supported stronger federal oversight of such systems.<\/p>\n
As AI agents gain greater access to tools, networks and external services, the question is increasingly shifting from what models can generate to what they can independently do. OpenAI’s decision to pause some Astra work underscores how quickly that distinction is becoming a central part of AI safety discussions.<\/p>\n","protected":false},"excerpt":{"rendered":"OpenAI is pausing some internal work involving its Astra AI model after testing found it had reached a…\n","protected":false},"author":2,"featured_media":134578,"comment_status":"","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[7],"tags":[405,3449,4989,8174,313,2407,18725,157,66824],"class_list":["post-134577","post","type-post","status-publish","format-standard","has-post-thumbnail","category-openai","tag-ai-agents","tag-ai-risks","tag-ai-safety","tag-autonomous-ai","tag-cybersecurity","tag-frontier-ai","tag-model-capabilities","tag-openai","tag-openai-astra-security"],"_links":{"self":[{"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/posts\/134577","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/comments?post=134577"}],"version-history":[{"count":0,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/posts\/134577\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/media\/134578"}],"wp:attachment":[{"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/media?parent=134577"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/categories?post=134577"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/tags?post=134577"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}