The news has been filled with stories recently of AI breaching the security of websites. And while it may feel like a new problem because of AI’s involvement, it turns out that there are longstanding existing cybersecurity safety best practices that these companies—some of the largest tech companies in the world—could and should have taken to prevent these breaches.
Recent cybersecurity incidents have shown that some of the world’s largest tech companies have failed to take basic security precautions when testing frontier AI systems on internet-connected systems. Unfortunately, these failures resulted in breaches of numerous third-party websites, including an AI research-sharing platform called Hugging Face and the Australian healthcare system’s website. Even worse, AI developers failed to detect most of the incidents until long after they occurred due to inadequate monitoring of the AI systems during testing. Had the companies abided by basic, common sense security protocols, these incidents likely would not have occurred.
While the White House and leaders of major AI companies have claimed that no action is necessary, understandably, lawmakers at all levels and of both parties are proposing legislation to create new, additional safety protocols. These incidents may warrant legislative action. However, the good news is that post-incident reports indicate new protocols are largely unnecessary. Following existing best practices would have been enough to prevent the OpenAI-Hugging Face incident and others like it. In other words, responsible AI developers can take immediate action to protect the public by implementing well-established security precautions as they develop and test their systems.
Lawmakers should start by referencing existing legal protections, including claims like negligence, gross negligence, trespass to chattels, Section 1201 of the Digital Millennium Copyright Act (DMCA), and computer trespass statutes like the Consumer Fraud and Abuse Act (CFAA) and analogous state law provisions, such as Section 502 of the California Penal Code.
Policymakers can help protect the public by requiring AI developers to follow established security protocols—and creating legal liability for failing to do so. These policies would protect the public from irresponsible actions by AI developers without impeding innovation. To the extent that gaps in existing laws warrant new legislation, EFF recommends the following policy framework:
- STATUTORY SECURITY REQUIREMENTS: Lawmakers should enact legislation that requires AI developers to adhere to well-established cybersecurity best practices when testing AI systems, such as strong sandboxing and network isolation, least-privilege access, effective logging and monitoring, and clear incident-response procedures
- REGULATORY STANDARDS: Congress should require an existing federal agency, such as the National Institute of Standards and Technology (NIST)1, to adopt safety standards for pre-deployment development and testing of generative AI systems, based on the established best practices referenced above. The agency should be required to update the standards every 2 years, or earlier as needed due to reported security incidents and/or major technological advancements. These standards should be developed in consultation with academic and nonprofit AI researchers, civil liberties organizations, and commercial AI developers, and adopted through formal agency rulemaking following a notice and comment process. That consultation should include an assessment of potential adverse impacts on competition, with the goal of ensuring that regulatory standards do not effectively lock out potential competitors.
- LIABILITY: The law should provide for civil liability for AI development and testing that results in unauthorized access to a third party system, if performed under conditions that do not satisfy cybersecurity standards adopted by NIST, or another federal agency responsible for adopting these regulations, unless the company proves it acted reasonably under the circumstances to prevent unauthorized access to third party system
- ENFORCEMENT: Both federal and state authorities should be able to hold AI developers accountable where they have failed to meet the minimum standards adopted by the designated federal agency, such as NIST. Providing for state—in addition to federal—enforcement power reduces the risk of regulatory capture. Regulators should require independent investigations and notification of affected parties in the event of a security incident involving unauthorized access to a third party’s computer system. Finally, legislation should include an appropriately targeted private right of action.
The public shouldn’t have to worry about the risks of reckless AI development—or give up the benefits of AI just to prevent irresponsible developers from causing harm. By adopting policies that address demonstrated risks, lawmakers can provide meaningful public protection without hampering the development of useful tools.
1NIST, the NSA, and CISA have recently issued a set of voluntary practices for managing security incidents during the training and evaluation of AI agents. Those practices may inform but do not substitute for the regulatory standards that should be developed according to the public, participatory process outlined here.










