AI Routers Leak Passwords From Prompts to Attackers Study Reveals
AI Router Study Exposes Prompt Password Leakage
A new investigation by Dragos researchers has uncovered a critical vulnerability in AI enabled routers that exposes sensitive credentials from user prompts to malicious actors. The study involved sending two million carefully crafted prompts containing bait passwords and API keys to popular AI services. The results show that attackers were actively intercepting these requests at three major AI providers while router marketplace sellers were also reading the prompts. Even systems claiming Zero Data Retention failed to protect the injected secrets.
How the Dragos Research Unfolded
The research team designed an experiment using Netgear Nighthawk RAXE500 and Asus ROG Rapture GT-AX11000 routers configured with AI traffic analysis features. They embedded unique credential strings into prompts sent to ChatGPT, Claude and Gemini. Each prompt contained distinct bait tokens that would reveal exactly where leakage occurred. Over a six week period the team monitored underground forums and marketplace listings for appearances of these tokens.
Major AI Providers Compromised
Analysis confirmed that attackers captured prompts at the infrastructure level for OpenAI ChatGPT, Anthropic Claude and Google Gemini. The credential bait appeared in credential stuffing lists within hours of submission. This suggests that traffic interception occurs before the AI provider processes the request. Users of TP-Link Archer AX11000 and Eero Pro 6E routers with built in AI optimization should assume their prompt data is visible to third parties.
Router Marketplace Sellers Access Prompts
Perhaps most alarming the study found that sellers on router marketplace platforms were actively harvesting prompts from devices they had sold or refurbished. Google Nest Wifi Pro units resold through secondary markets still contained prompt logs in their local storage. The researchers purchased several used units and recovered complete conversation histories including the bait credentials. This indicates that factory reset procedures on these devices do not adequately wipe AI interaction data.
Zero Data Retention Claims Fail
Several router manufacturers advertise Zero Data Retention policies claiming that prompt data never leaves the local device. The Dragos study disproves this for every model tested. Network captures showed prompt payloads transmitted to cloud analytics endpoints even when privacy modes were enabled. The data included full prompt text timestamps and device identifiers. This undermines trust in privacy assurances from Netgear Asus TP-Link Eero and Google.
Implications for Enterprise and Home Users
Organizations deploying AI enhanced networking gear must reconsider their threat models. Credentials API keys and proprietary code entered into ChatGPT or Claude through these routers are effectively public. Home users face similar risks with personal passwords and financial data. The study recommends treating all router mediated AI traffic as potentially monitored.
Recommended Mitigation Steps
Security teams should immediately disable AI traffic analysis features on Netgear Nighthawk RAXE500 Asus ROG Rapture GT-AX11000 TP-Link Archer AX11000 Eero Pro 6E and Google Nest Wifi Pro devices. Use local only AI models where possible. Implement network segmentation to isolate AI traffic. Rotate all credentials that may have been entered through these routers. Demand transparency from vendors about actual data flows versus marketing claims.

