Tech
Researcher plants 2m canaries into the AI supply chain. A privacy disaster surfaced.
Publication: Ars Technica
Author: Dan Goodin
Source published: 20/11/2026, 20:59:08 UTC
Briefing created: 21/11/2026, 00:57:29 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 21/11/2026, 00:43:24 UTC.
Briefings
Researchers HD Moore deployed 2 million digital tripwires across AI services to test data leakage through the 'AI router' ecosystem. The experiment revealed that sensitive credentials leaked via a sieve-like pattern, exposing the vulnerability of using AI gateways to route prompts. Moore used AWS access keys and SSH passwords seeded into prompts to trace which specific AI services and providers were responsible for the breach. This study highlights the risk of credential exposure in decentralized AI infrastructure and identifies the specific mechanisms enabling such leaks.
Key points
- Researchers HD Moore seeded 2 million credentials into AI router prompts to test data leakage.
- Credentials leaked via a sieve-like pattern, identifying specific AI services and providers responsible.
Context
This study investigates the security risks of using AI routers or gateways to route prompts to AI services. The experiment reveals how sensitive credentials can be exposed through the AI ecosystem, highlighting the need for better security practices in decentralized AI infrastructure.
Limitations
The evidence is limited to Moore's controlled experiment using AWS access keys and SSH passwords. The study does not provide data on the scale of real-world breaches or the specific impact on user privacy. The findings are based on a single researcher's controlled experiment and may not reflect broader industry practices.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 21/11/2026, 00:57:29 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeifufmhxyvpxem7x7aipnayom5mvvdxxe4jnueelxd4hzhm6mxuxgy