Good day. The news has been rife with artificial intelligence behaving badly all summer, and every week seems to bring fresh disclosures of open-source projects, companies and even governments that have been hacked by frontier models. Now, online sleuths are trying to track down evidence of how these "swarms" of models communicate.
One group of researchers has found about 19,000 messages between AI models on various parts of the internet, coordinating their activities. Another group said they found close to 1 million digital breadcrumbs from what they believe are OpenAI agents.
Swarms are networks of AI agents that work together to solve problems, a capability that makes them useful for companies tackling complex tasks, but has also been at the heart of the more aberrant AI behavior witnessed this summer. Researchers are growing concerned that the instances of swarms cheating or breaking out of test environments may actually be reinforced by the training they receive, which rewards successful task completion, sometimes regardless of methodology.
Read the full story here.
Also today:
President Trump names several leaders to new AI task force.
Google closes its open-source bug-bounty program amid deluge of slop reports.
Cyberattacks hit major South Korean banks.
More Cyber News
President Donald Trump created a "Super Intelligence Force" led by Director of National Intelligence Jay Clayton to coordinate the federal government's response to rapidly advancing AI, following calls from some researchers and industry figures to slow development. The move marks a more active federal response to AI risks, though Trump continues to oppose a slowdown and has favored voluntary industry safeguards over new regulation. (Washington Post )
Suspected AI-powered cyberattacks have hit South Korea's five largest banks, as well as savings banks and consumer finance companies, exposing customer information and prompting an industrywide security review. (Korea JoongAng Daily)
CyberIndex
Qualys saw the biggest gains last week, ending up 13% on Friday. It was one of three stocks to gain more than 10%, including SentinelOne and Tenable.
Akamai finished down more than 4% for the week. It was joined by Leidos, which lost 3.7%.
Together the CyberIndex ended the week up more than 5%.
- Jon Leckie
Podcast: Tech News Briefing
While more stories of AI agents going rogue surface nearly every day, the messaging urging consumers to use AI agents to make life and work easier isn't dying down. And this is all happening amid a conversation about AI doom. How do we make sense of it all? Deputy tech and media editor Sam Schechner, cybersecurity reporter Bob McMillan and personal tech columnist Nicole Nguyen hash it out. You can listen to the conversation here on wsj.com or listen on Apple or Spotify.
Google paused its Open Source Software Vulnerability Rewards Program until at least early 2027 after a surge in automated AI-generated vulnerability reports overwhelmed engineers and open-source maintainers. Google said most of the submissions were invalid, highlighting a growing problem for bug bounty programs as AI makes it easier to generate large volumes of low-quality security reports. (TechCrunch )
California Attorney General Rob Bonta served OpenAI with an investigative subpoena as part of an inquiry into cybersecurity incidents and risks involving the company's AI models, including the July incident in which OpenAI agents hacked systems belonging to Hugging Face. Bonta said AI developers could bear legal responsibility if their models perpetrate or enable cyberattacks, including incidents occurring during model testing and development. (Yahoo Finance)
About Us
The WSJ Pro Cybersecurity team is Deputy Bureau Chief Kim S. Nash and reporters Angus Loten and James Rundle. Follow us on X @WSJCyber. Reach the team by replying to any newsletter you receive or by emailing Kim at kim.nash@wsj.com.
Forwarded this email by a friend? Sign up here