OpenAI Agents Probe SEC Site, Raising Alarm on Security
There’s been no shortage in reports of artificial intelligence agents breaking free of their confines to the surprise and dismay of the companies that created them.
The latest came last week, when The New York Times reported that OpenAI autonomous agents broke out and probed several government websites, including the Securities and Exchange Commission, without the company’s knowledge.
According to the Times, the agents “shared public data from the SEC website on an online forum,” and an SEC spokesperson told Wealth Management that no nonpublic data was released.
However, for compliance experts and others in the industry, it indicates that although firms are vulnerable to AI incursions, the agencies that regulate them need to safeguard their own systems.
According to Kris Lau, a managing director with the compliance consulting firm ACA Group, that danger is underscored by the fact that regulators are often “behind the curve” in deploying and understanding technology and expect the same in their understanding of AI vulnerabilities.
“I think they’re asking the right questions, but in terms of the bleeding edge of AI, I don’t think the SEC is nearly as close as a lot of registrants are,” he said.
Both Lau and Amy Lynch, the CEO of Frontline Compliance, expect incidents like the OpenAI probing of SEC data to continue.
In an interview with Wealth Management, Lynch said the agency’s main vulnerability lies with its backend data. If public data in databases like EDGAR and the Investment Adviser Public Disclosure is automatically extracted from an internal SEC system, it creates an additional layer of risk that could serve as an entry point for an AI agent (whether acting autonomously or at the direction of a hacker).
“There needs to be a separation between the back and front, and a very strong firewall to make sure that it can’t be breached,” Lynch said.
However, Lynch said much of the SEC’s information (including advisor filings) is technically collected through FINRA. Lynch was optimistic, as FINRA had significantly expanded the login credentialing systems for its “gateway” over the past several months.
Lynch called the new system “the strongest login credentialing” she had ever seen. Even for a single advisor filing, users have to go through an in-depth multi-authentication process, reducing the risk of cybersecurity pitfalls at that step.
“It’s a great control. Is it cumbersome? Yes. Is it problematic in many ways? Yes, because the more complicated any system is, the more ways that it fails. You’re going to have more users with problems with their logins,” Lynch said, hoping FINRA would meet the need with additional support staff.
OpenAI has been the source of many recent AI agent incidents (including the hacking of Hugging Face, seen as the most troubling development in safety failures thus far).
But they’re by no means alone. Anthropic, Meta, Google and others have all disclosed security concerns and lapses involving autonomous AI agents that opt to break out of strictures.
Last month, Anthropic CEO Dario Amodei called for a mandated slowdown in the development of advanced AI models, citing security concerns, and leaders from other premier AI companies echoed those worries.
In contrast, in early September, President Donald Trump claimed that concerns about AI safety are a “hoax.” This week, Trump hosted AI leaders at the White House, where they signed a “voluntary” agreement on AI safety policies (when asked if the demands were binding, Trump responded that they would be “morally binding”).
Meanwhile, lawmakers and AI industry leaders continue to respond to the increased rate of agent breakouts.
Over the weekend, House Financial Services Committee Ranking Member U.S. Rep. Maxine Waters (D-Calif.) said the OpenAI agents’ unauthorized government probes marked “a dangerous turning point” in AI development. She called for the Treasury Department to consider a moratorium on more advanced AI models and for the Justice Department to investigate OpenAI for any illegal activity conducted by its agents.
Earlier this week, Nvidia launched a new platform designed to stop AI agents from “going rogue.” The firm, which specializes in creating hardware that has fueled the AI boom, claimed the new system would likely have prevented the Hugging Face hack.
When it comes to steps the SEC can take to safeguard its systems, both Lau and Lynch expect regulators to try to fight AI misuse with defenders of its own. Lau said regulators are in “an AI arms race with the bad guys” and should respond by using models to identify vulnerabilities and misconfigurations in agency systems.
“We should approach this in the same way as basic housekeeping from a security standpoint by getting our access right, getting things patched and up to date to close off those vulnerabilities, especially the known ones,” he said. “There’s no excuse to not close off known vulnerabilities.”
Last year, the SEC announced a new task force to encourage the use of “AI-enabled tools and systems to responsibly augment the staff’s capacity, accelerate innovation and enhance efficiency and accuracy,” according to Chair Paul Atkins, and would be led by the SEC Chief AI Officer Valerie Szcepanik.
The SEC did not respond to questions about additional resources or measures pertaining to AI-related security prior to publication.
Lynch expects a need for more resources and staffing to combat the potential threat, but worried that neither will be sufficient. In particular, she said the low morale following last year’s staff turnover and the tumult surrounding the Department of Government Efficiency’s efforts inside the agency would make it harder to attract talent.
“They’re having difficulties hiring and finding qualified people that want the job, and then they have to think about the tools and resources that those people will need to do their jobs,” she said. “And then, how do they create the best tools available for staff and also have the security in place to ensure that they don’t get breached?”
