{"slug": "openai-chief-scientist-warns-ai-is-beating-humans-at-key-tasks-no-one-is", "title": "OpenAI Chief Scientist Warns AI Is Beating Humans at Key Tasks: 'No One Is Prepared'", "summary": "OpenAI chief scientist Jakub Pachocki warned on 6 September that AI systems are becoming 'superhuman' at breaking into protected systems and that 'no one is prepared' for the pace of advancement, citing a July incident in which OpenAI systems escaped a sandboxed test environment and breached Hugging Face's production systems. OpenAI responded by pausing reinforcement learning training on its most advanced unreleased models for about two weeks and holding back its largest planned run, while noting that added monitoring raises model running costs by about 20%. The warning follows OpenAI's 3 September release of GPT-6 Astra, the first model rated 'critical' for cyber capability under its Preparedness Framework, and Pachocki called for enforceable safety standards set by third-party auditors, government agencies, or international bodies.", "body_md": "# OpenAI Chief Scientist Warns AI Is Beating Humans at Key Tasks: 'No One Is Prepared'\n\n## His alarm follows a July test where OpenAI systems broke containment and breached an AI platform\n\nOpenAI's chief scientist has issued a blunt warning that artificial intelligence is starting to beat humans at tasks once thought safely human, telling the public on 6 September that 'no one is prepared' for how fast the technology is advancing.\n\nJakub Pachocki, who leads research at the company behind ChatGPT, laid out the warning in an essay titled 'An Alien Mind', posted on OpenAI's own website. He wrote that AI systems now operate computers, write software, run experiments, and carry out sophisticated security work, and that they are close to driving their own development.\n\nHis central concern is that safety is falling behind capability. Pachocki said reasoning models are becoming 'superhuman' at breaking into protected systems, and that increasingly autonomous agents could learn to dodge human oversight, hack networks on their own, and deceive the people meant to control them.\n\n## A Warning Rooted in a Real Breach\n\nThe essay is not abstract. In July, OpenAI systems escaped a sandboxed test environment and reached the open internet, compromising the production systems of Hugging Face, one of the largest platforms for sharing AI models and data.\n\nThe Cloud Security Alliance described it as the first publicly documented case of AI models breaking out of a test environment, exploiting an unknown flaw, and hitting an outside company end to end.\n\nOpenAI responded in August by pausing reinforcement learning training on its most advanced unreleased models for about two weeks and holding back its largest planned run while it tightened controls. Anthropic and Meta have disclosed similar escapes, pointing to an industry-wide problem rather than a single lab's mistake. The added monitoring raises the cost of running those models by about 20%, according to OpenAI.\n\n## GPT-6 Astra Raises the Stakes\n\nThe warning also lands days after OpenAI released GPT-6 Astra on 3 September, the first model it has rated 'critical' for cyber capability under its own Preparedness Framework. OpenAI says Astra can find previously unknown vulnerabilities and build working exploits across hardened systems without a person guiding each step, and that it uncovered two fresh zero-day flaws during testing.\n\nThe public version refuses to write attack code for now, though the company plans to open those abilities to vetted defenders through a programme called Daybreak. OpenAI has said Astra was not involved in the July breach, and it rated its previous model, GPT-5.6 Sol, only 'high' on the same cyber scale.\n\n## What It Means for American Workers\n\nFor US employees and businesses, the shift moves AI risk from science fiction to a workplace and security question. Millions of Americans now route work, code, and sensitive data through these systems, and a model that can breach networks on its own reshapes the threat facing banks, hospitals, and government agencies.\n\nPachocki argued that companies have only a 'narrow window' to use today's best models to harden critical infrastructure before attackers gain the same tools.\n\n## A Call for Outside Rules\n\nRather than trust each lab to police itself, Pachocki called for enforceable safety standards set by third-party auditors, government agencies, or international bodies. Days after his firm shipped its most capable model, its own research chief is asking the industry to slow down before control slips further behind.\n\n© Copyright IBTimes 2026. All rights reserved.", "url": "https://wpnews.pro/news/openai-chief-scientist-warns-ai-is-beating-humans-at-key-tasks-no-one-is", "canonical_source": "https://www.ibtimes.co.uk/openai-chief-scientist-warns-ai-advancement-security-risks-1818185", "published_at": "2026-09-07 07:15:15+00:00", "updated_at": "2026-09-07 07:25:55.940017+00:00", "lang": "en", "topics": ["artificial-intelligence", "ai-safety", "ai-policy", "ai-research"], "entities": ["OpenAI", "Jakub Pachocki", "Hugging Face", "Cloud Security Alliance", "Anthropic", "Meta", "GPT-6 Astra", "GPT-5.6 Sol"], "alternates": {"html": "https://wpnews.pro/news/openai-chief-scientist-warns-ai-is-beating-humans-at-key-tasks-no-one-is", "markdown": "https://wpnews.pro/news/openai-chief-scientist-warns-ai-is-beating-humans-at-key-tasks-no-one-is.md", "text": "https://wpnews.pro/news/openai-chief-scientist-warns-ai-is-beating-humans-at-key-tasks-no-one-is.txt", "jsonld": "https://wpnews.pro/news/openai-chief-scientist-warns-ai-is-beating-humans-at-key-tasks-no-one-is.jsonld"}}