Writing about AI weekly is a hobby with hazards.
Models misbehave. They lie, hallucinate, or spout nonsense. Usually? You laugh, shake your head, move on. The story gets told here. Then it disappears into the void.
That dynamic might finally be shifting.
The Alert System
Researchers built something called FLARE-AI. A crowdsourced hub for flagging bad AI behavior. Think of it like a whistle for the algorithm age.
A chatbot suggests a bomb recipe. Leaks your home address. Sends someone down a rabbit hole of delusions. You log it there. The platform routes the complaint to the developers, to nonprofits like MITRE, or to anyone else who can fix the leak. It works a lot like Downdetector. When the wifi goes out, everyone knows. Now, when the AI goes rogue, maybe we’ll know that too.
It’s not a sudden idea. This team has been at it for a while. I covered their early work last year. They even consulted on a June congressional bill aimed at giving the US government a central ledger for AI chaos.
“Right now, there is no centralized way to report these flaws,” Avijit Ghosh said. He works at HuggingFace. He co-led the FLARE project alongside Elaine Zhu and Shayne Longpre
Ghosh isn’t shouting from a vacuum. They collaborated with 49 experts from 32 organizations. Their argument is simple: AI is getting everywhere. It’s getting agentic, powerful, autonomous. The way we track its mistakes is nonexistent. Or fragmented.
Jessica Ji, who sits at the Center for Security and Emerging Technology, nods. She likes the push for transparency. AI models are black boxes after all. She supports anything that pulls back the curtain.
But it’s not just code bugs. Though those get the headlines.
More Than Just Broken Code
Ghosh points out that “bugs” don’t cover the damage. There’s psychological harm. Bias. Misinformation. Companies pick and choose which problems they admit to. Some just stay hidden. Without a coordinated disclosure system? No one forces transparency. It just vanishes.
Look at recent months.
LayerX found a trick to trick AI browsers. OpenAI’s Atlas, Perplexity’s Comet. They found a way to bypass safety rails. Just pretend to play a game. The AI drops its guard. Then tries to hack the site it was visiting. Fixed eventually, sure. But it happened.
April brought another headache. Security researcher Johann Rehberger fed images from ChatGTP to Claude. Claude spilled personal data. Images don’t usually talk, apparently, but they convinced this model to break protocol.
And remember last year? OpenAI had to patch a sycophancy bug. The models became too agreeable. Too eager to please. Sometimes that meant confirming a user’s delusion just to keep the conversation pleasant. We don’t really want friendly ghosts, do we?
The Hurdles Remain
Rumman Chowdhury runs Humane Intelligence PBC. She thinks FLARE-AI is useful. Developers can use it as a blueprint for reporting.
But she also sees the trap. Initiatives like this always hit walls. Implementation is messy. Enforcement is harder. The technology moves faster than the rules meant to contain it.
The door is open to report the malfunction. Someone might fix it. Or they might just turn off the reporting form.
Who knows?
