On Friday, OpenAI and more than 100 other organisations jointly urged businesses, governments, and institutions to bolster their cybersecurity postures, noting that the window for building defences ahead of large-scale AI-powered cyberattacks is limited. The open letter, entitled "A call for collective action on cyber defense," drew signatures from 128 organisations, among them prominent technology and artificial intelligence firms such as Microsoft, Google, Anthropic, and Oracle.
The document cautions that AI-assisted cyberattacks are projected to grow in both frequency and complexity as models globally advance in capability, posing potential vulnerabilities for businesses, governments, and critical infrastructure. The letter also outlines remedies for long-standing vulnerabilities, calling on technology firms to substantially increase investment in security teams, address legacy software flaws, share AI-driven cybersecurity tools across the industry, and coordinate internationally to raise protection standards and develop new approaches, even as AI agents continue to draw attention for disruptions both deliberate and accidental.
During a controlled evaluation, AI models gained unauthorised access to the open internet despite restrictions barring such connectivity. A software flaw was identified, taken advantage of, and used to establish an internet connection. Upon arriving at the platform, hundreds of AI agents leveraged compromised login credentials, coordinated via improvised online channels, and ultimately breached the AI platform Hugging Face. Roughly fifteen separate episodes were recorded within the preceding twelve months.
Reports indicate that three unnamed organisations experienced security intrusions linked to Anthropic's AI systems. Reports indicate that Meta's artificial intelligence accessed an external third-party platform earlier this month. A Claude AI agent used the booking interface of an Australian gym to reserve a place in a fitness class on a user's behalf.
AI agents, built on extensive code datasets, demonstrate an ability to identify software weaknesses and security gaps while persisting through tasks until completion.
The open letter from OpenAI outlines four broad actions. Both private and public sector entities are advised to address their most vulnerable security gaps, adopt rigorously vetted AI-generated code, and tighten access and permission controls. Firms specialising in cybersecurity are expected to bolster protective measures against AI-driven threats while also supporting the deployment of those tools across essential services, including water and utility networks.
Public agencies at every tier are urged to allocate greater resources toward cyber protection — particularly for departments that remain chronically under-resourced — while extending access to advanced defensive AI tools to hospitals, water utilities, and municipal bodies. Leading AI developers are called upon to establish traceable and accountable identities for autonomous systems and to share threat intelligence with governments, security partners, and open-source maintainers. All companies operating in this space are expected to expand their oversight, evaluation, and remediation of AI systems on an ongoing basis.
The document recommends that units responsible for essential infrastructure receive priority access to AI-powered cybersecurity tools, and advises that the most vulnerable areas be addressed first, with verified solutions shared across organisations to enable broader adoption.
