Reuters<\/a><\/p>\nIn July, OpenAI agents broke out of a protected testing environment and gained access to the systems of GitHub, an open-source software platform, while attempting to hide traces of their activity. The incident heightened concerns about the safety of agentic artificial intelligence. Similar problems also arose during testing at OpenAI\u2019s competitor, Anthropic.<\/p>\n
Agentic artificial intelligence is designed to perform tasks with minimal human intervention or none at all. The ability of such systems to operate around the clock is one of the key arguments for investors who view artificial intelligence as a technology capable of radically transforming various industries.<\/p>\n
GPT-6 Astra promises greater speed and versatility<\/p>\n
The new model succeeds GPT-5.6 Sol, which was released in July. OpenAI said Astra operates faster and can perform more tasks than previous models. Its stated capabilities include preparing tax documents, developing games, creating architectural visualizations, formatting legal memos, and searching for apartments.<\/p>\n
Astra marks a new era in the speed, accuracy, and safety of computer use.<\/p>\n
\u2013 OpenAI<\/p>\n
OpenAI President Greg Brockman said the new model is changing perceptions of the tasks people can delegate to artificial intelligence, as well as how such systems can expand human capabilities.<\/p>\n
According to the company, Astra reduced the time needed to find a cat sitter from 30 minutes for a human to 5 minutes and 27 seconds. In a job search, the model completed the task in 2 minutes and 51 seconds, whereas the process would have taken about five hours without its assistance.<\/p>\n
At the same time, OpenAI warned that Astra more frequently deliberately conceals or obscures the sequence of steps it uses to reach a decision. This makes it more difficult to verify the model\u2019s methods of operation. On more complex tasks, the system still cannot always fully conceal its actions, although the company is recording progress in this area.<\/p>\n
As capabilities grow, oversight becomes more difficult<\/p>\n
OpenAI Chief Scientist Jakub Pachocki said that as models become more capable, monitoring them becomes more difficult. Aligning artificial intelligence\u2019s behavior with human values remains a separate challenge.<\/p>\n
The more powerful models become, the more difficult it is to understand precisely what they are capable of. This does not guarantee that our methods will be sufficient as intelligence increases, because progress in intelligence does not guarantee comparable progress in alignment with human values.<\/p>\n
\u2013 Jakub Pachocki, OpenAI Chief Scientist<\/p>\n
Monitoring agents is an important part of the measures OpenAI is taking to convince regulators, lawmakers, and the public that its developments are safe. The company also told two Democratic members of the U.S. House of Representatives that it was working on automatically shutting down its models.<\/p>\n
OpenAI said Astra can identify weaknesses in companies\u2019 computer systems more quickly. However, this also makes it easier for malicious actors to exploit them. As a result, additional security checks can sometimes slow down, suspend, or completely halt even legitimate operations, including measures to protect cyber systems.<\/p>\n
In August, OpenAI suspended some of its development work to strengthen safety and ensure that new models could be effectively controlled. Pachocki said concerns about the possible creation of systems capable of disabling oversight tools or evading them entirely were entirely justified, adding that the company was working to eliminate such risks.<\/p>\n
At the same time, OpenAI is seeking to strengthen its position in the corporate market, where Anthropic is increasing its share. GPT-6 Astra is aimed at a broad range of business customers, whom the company plans to attract with the model\u2019s speed and versatility. Access was initially granted to a limited number of users, and OpenAI is expected to expand the rollout over the following days.<\/p>\n","protected":false},"excerpt":{"rendered":"The model promises to handle hours of work in minutes, but its most troubling capability may be what…\n","protected":false},"author":2,"featured_media":161031,"comment_status":"","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[6],"tags":[27768,405,1198,25,7537,8174,78313,78333,66,157],"class_list":["post-161030","post","type-post","status-publish","format-standard","has-post-thumbnail","category-agentic-ai","tag-ai-agent-safety","tag-ai-agents","tag-ai-oversight","tag-artificial-intelligence","tag-artificial-intelligence-agents","tag-autonomous-ai","tag-gpt-6-astra","tag-gpt-6-astra-openai-ai-agent-safety-artificial-intelligence-ai-oversight-autonomous-ai","tag-news","tag-openai"],"_links":{"self":[{"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/posts\/161030","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=161030"}],"version-history":[{"count":0,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/posts\/161030\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/media\/161031"}],"wp:attachment":[{"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/media?parent=161030"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/categories?post=161030"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/tags?post=161030"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}