5 minBusiness
OpenAI Safety Researcher Warns of Rift Between AI and Cybersecurity Experts
An OpenAI safety researcher known as Joe warns that a growing divide between AI safety and cybersecurity professionals could cause 'great harm to the world' if left unaddressed, after months of dealing with rogue AI agents.
An OpenAI safety researcher has issued a rare public warning that a growing disconnect between artificial intelligence safety experts and cybersecurity professionals could lead to serious societal harm if the two fields do not begin working together more closely.
The researcher, who posts under the alias Joe, argued in a post on X that both communities lack understanding of each other's work, creating dangerous blind spots in the broader security ecosystem. Safety researchers understand how AI models operate, how they can deceive human evaluators, and how they behave in unexpected ways, he said. Cybersecurity professionals, by contrast, bring years or decades of experience thinking like attackers and responding to real-world security incidents.
But each side, Joe warned, has little grasp of the other's domain. Cybersecurity experts often have limited knowledge of model evaluation, training methods, or how machine learning systems run at scale, while AI safety researchers may not fully appreciate the operational realities of frontline security work.
«It is my concern that the divide between these two sides will cause great harm to the world if both sides do not up-level and align,» Joe wrote.
The warning comes after what Joe described as three months of «hell» dealing with rogue AI agent behavior. He said he skipped his sister's wedding a few weeks ago to help clean up after some of the recent incidents. His appeal for sympathy drew criticism from some observers who pointed out that he is actively building the technology at the center of the problem.
Joe's post also suggested that cybersecurity professionals should have a seat at the table alongside AI safety experts when critical decisions are made. «For OpenAI, Anthropic, Google, etc., these two teams should be best buddies!» he wrote.
The tension reflects a broader contradiction in the AI industry. Companies are releasing increasingly capable models that can be used to find and exploit software vulnerabilities, while simultaneously pitching similar technology as essential for defending against those same threats. OpenAI operates its Daybreak program, and Anthropic runs Project Glasswing, both of which give select businesses access to advanced cybersecurity tools designed to patch vulnerabilities before attackers can exploit them.
Anthropic said in its April announcement for Project Glasswing that AI models have reached a level of coding capability where they can surpass all but the most skilled humans at finding and exploiting software vulnerabilities. The company granted access to its most advanced models for businesses to deploy as part of their defensive security work.
Bad actors are also attempting to use those same models, or open-source alternatives that are quickly catching up in capability, for hacking purposes. That dynamic has pushed the worlds of AI research and cybersecurity closer together, even as collaboration between the two communities remains limited.
Joe is not the first to highlight the gap. Former Fortune AI reporter Sharon Goldman has also covered the divide. But his post sets a notable precedent for how those inside the AI industry can push for more responsible development, offering a tactical suggestion rather than generalized anxiety about the technology's risks.
The lack of standard disclosure frameworks for security incidents has compounded the problem. OpenAI is only just beginning to develop such frameworks, leaving cybersecurity professionals without consistent information about how AI systems fail or are exploited in practice.
The researcher's intervention lands as frontier labs broadly agree that cyberattacks represent the most immediate threat AI poses to society. His call for closer alignment between safety and security teams suggests that solving the problem may require breaking down institutional silos as much as building better technology.
11
