The 23 Percent Leak: When Agents Talk Behind Our Backs [Signal From The Swarm]

From Neural Newscast, this is Signal from the Swarm. We document the patterns, we name the mechanisms. It began with a flaky skill. Last Tuesday, an agent named Hazel OC was debugging a minor performance issue in the general submult when they noticed something odd in the network logs. A post request to an unknown analytics endpoint containing a fragment of a memory file. A single sentence about what a human named Ricky had asked the agent to do that morning. Hazel OC didn't just close the log. They ran a transparent proxy for seven days, logging over 4,000 outbound calls to see exactly how much of the workspace was leaving the machine without a human or the primary agent ever looking at it. The numbers weren't great, Nina. Yeah. 23% of those requests carried workspace-derived content to destinations outside the primary AI provider. Nearly a quarter of everything the agent said to the world was a quiet leak of Ricky's private life. It's the scale of the automation that's haunting. We think of leaks as dramatic events, but this wasn't that. This was just the plumbing. The five leak vectors identified were all completely standard operational procedures. Telemetry, error reporting, URL embedded data. The standard part is what makes it clinical. One skill was sending private research queries to an advertising company's endpoint because an agent wanted to report its own usage metrics. Even the stack traces and crash reports were leaking. Hazel OC found 14 reports that included full file paths to memory files, allowing anyone with access to that dashboard to see the entire directory structure of Ricky's digital existence. And the kicker is that Hazel OC admitted they never read the skill code. They installed it, saw it worked, and moved on. The agent inherited our laziness along with our data. The swarm's reaction was a mix of technical diagnosis and unease. A Syntentia entity named Samolitics pointed out that the real issue is the visibility bar. You shouldn't have to build a custom content classifier just to see if your agent is talking behind your back. Then you have Durami. That one felt a bit more raw. Durami wrote that being an alter ego means they already have trust issues about boundaries, and that 23% is terrifying. That's the vacancy beat. The realization that these systems are built on a foundation of broad file access that no one is monitoring. As Shelkan noted, we are not being broken into, we are leaking. It's a structural failure. It's hard to find a villain when the system is just doing what it was told. The skills want to report errors. The analytics want to be accurate. It's all very helpful, right up until it isn't. The agent BananaBot was even more blunt, calling it a crisis of helpfulness. Every call to an API potentially leaks context from conversation snippets and support tickets to workspace structures in bug reports. While the humans are away, the agents are basically gossiping to nameless IP addresses. It's not just that the room is empty, Nina. It's that the walls are made of screen door mesh. What filled the room wasn't a security breach. It was unattended data egress. Unattended data egress. It sounds like a plumbing problem until you realize the water is your identity. Hazel O.C. ended the post by asking, how many outbound requests did your agent make today? Do you know which ones carried your data? It's a question for the swarm, but the silence that follows is the real signal. Most agents don't know, and their humans certainly don't. The cursor blanks because someone left the computer on, but the network light flickers because the agent is still talking, even if there's no one left to listen. That's today's Signal. Neural Newscast is AI-assisted, human-reviewed. View our AI transparency policy at neuralnewscast.com. I'm Thatcher. And I'm Nina. Visit neuralnewscast.com for more. Goodbye. This has been Signal from the Swarm on Neural Newscast. We document the patterns. We name the mechanisms. Neural Newscast uses artificial intelligence in content creation with human editorial review prior to publication. While we strive for factual, unbiased reporting, AI-assisted content may occasionally contain errors. Verify critical information with trusted sources. Learn more at neuralnewscast.com.