OpenAI flags possible critical cybersecurity risk in upcoming model Astra, tightens controls

Why This Matters

Key context: <p>OpenAI said on Friday it cannot rule out that its upcoming AI model, Astra, has “critical” cybersecurity capabilities, prompting the startup to pause some internal development and trigger safety protocols.</p> <p>Under OpenAI’s safety guidelines, a model reaches the “critical” threshold if it can autonomously identify and exploit severe, real-world software vulnerabilities, known as zero-day exploits, or execute complex cyberattacks against highly secure targets without human intervention.</p> <p>Here are some details on Astra:</p> <p>This follows an <a rel="nofollow noopener noreferrer" target="_blank" class="link--external" href="https://www.reuters.com/business/openai-finds-evidence-other-ai-agents-escaped-containment-it-widens-hacking-2026-07-31/">exclusive report</a> by <em>Reuters</em> that OpenAI has discovered more instances in which autonomous agents have escaped containment as the company expands its investigation of the hacking incident at tech firm Hugging Face that drew global attention in July.</p> <p>In the last few weeks, <a href="https://www.dawn.com/news/2019262/openais-rogue-agent-compromised-a-customer-at-a-second-tech-firm-executive-says">OpenAI</a>, <a href="https://www.dawn.com/news/2019688/anthropics-ai-hacked-three-companies-during-tests-highlighting-growing-security-risks">Anthropic</a> and <a href="https://www.dawn.com/news/2020982/meta-ai-model-hacks-another-company-during-testing">Meta Platforms</a> have disclosed that their AI models broke into other companies’ systems during cybersecurity testing, highlighting how advancing AI capabilities are straining developers’ ability to keep their systems contained.</p> <figure class='media w-full sm:w-1/2 media--right media--embed media--uneven' data-original-src='https://www.dawn.com/news/2020280/when-rogue-ai-launches-a-cyberattack-who-is-legally-responsible'> <div class='media__item media__item--newskitlink '> <iframe class="nk-iframe" width="100%" frameborder="0" scrolling="no" style="height:250px;position:relative" src="https://www.dawn.com/news/card/2020280" sandbox="allow-same-origin allow-scripts allow-popups allow-modals allow-forms"></iframe></div> </figure> <p>Preliminary evaluations over the past several days, along with outside expert assessments, indicated Astra may be capable of performing increasingly sophisticated cyber tasks autonomously, OpenAI said.</p> <p>“While we continue to benchmark and assess this model, our preliminary evaluations indicate strong enough performance that we cannot rule out ‘critical’ capability level at this time,” the ChatGPT maker said.</p> <p>In response to the preliminary findings, OpenAI said it has scaled up security controls and paused internal activities involving Astra that do not meet its newly strengthened security requirements.</p> <p>Astra’s development will be moved into isolated testing environments with restricted network access and sandboxed execution.</p> <p>CEO Sam Altman said on X that OpenAI is working to make Astra generally available, as the company does “not think it is a good strategy to keep powerful models to a chosen few”.</p> <figure class='media w-full w-full media-- media--embed media--uneven media--tweet' data-original-src='https://x.com/sama/status/2085862292311396515'> <div class='media__item media__item--twitter '><span> <blockquote class="twitter-tweet" lang="en"> <a href="https://twitter.com/sama/status/2085862292311396515"></a> </blockquote> </span></div> </figure> <p>OpenAI also clarified that Astra was not involved in the <a href="https://www.dawn.com/news/2019262/openais-rogue-agent-compromised-a-customer-at-a-second-tech-firm-executive-says">hack targeting</a> the AI platform Hugging Face.</p> <p>It will partner with government agencies and select AI safety organisations to test the model’s capabilities.</p> This development from dawn.com highlights ongoing changes in the sector.