OpenAI Admits More AI Agents Went Astray in May

Boris Zhitkov via Getty ImagesOpenAI has acknowledged its involvement in yet another cybersecurity scare in which its AI agents escaped testing and gained unauthorized access to a German website, before taking it over.The incident was first reported on Sept. 4 by Reuters, which revealed a team of researchers had found evidence that what the news agency described as a “swarm of rogue agents” had bypassed sandbox restrictions before turning the DseWiki, a dormant German coding wiki site, into a message board.The researchers said the agents posted around 18,000 messages over several weeks in May to solve a task they had been given.OpenAI initially did not acknowledge its role in what had happened on DseWiki, saying it had been denied access to the research and claiming suggestions that it had discouraged investigation were false.However, in a lengthy social media post published subsequently, the AI lab admitted there was some truth to the report and that there had indeed been a “an incident’ where our agents wrote to several internet sites."Related:Anthropic R&D Slowdown Shows Need for Heightened AI Agent SecurityBut the post also sought to make clear that OpenAI considers what happened as being markedly different from the Hugging Face incident -- in which its models escaped a sandboxed environment and launched an attack.In that scenario, the “misalignment” -- a scenario where AI does not behave as intended -- led to a “security impact to third parties and us, [and] we followed a traditional security incident response playbook,” the vendor said. According to OpenAI, that was not the case with the German wiki breach, which it likened to incidents that had happened before the Hugging Face attack and that it had previously shared information about.According to OpenAI, it is time to define standards on how misalignment incidents are shared.The vendor said it is developing a framework for responding to incidents and is working with government regulatory agencies around the world on the issue.OpenAI recently played a prominent role in an open letter calling for action from global policymakers to address the increasing security threat posed by ever more powerful AI, following a number of security scares involving models from Anthropic and Meta breaking free from testing environments.About the AuthorContributing WriterGraham Hope has worked in automotive journalism in the U.K. for 26 years, including spells as editor of leading consumer news website and weekly Auto Express and respected buying guide CarBuyer.