Dev.to Security πŸ” Cybersecurity πŸ‘ 0 πŸ“– 4 min read

OpenRouter Routed Some Chatbot Prompts to a Hacked Server for 18 Days. Were Your Passwords in Them?

On 10 November 2026, Phala published a preliminary incident report. Phala is a company that OpenRouter sometimes sends chatbot requests to. The report says an attacker took over a server that was running the DeepSeek V4

On 10 November 2026, Phala published a preliminary incident report. Phala is a company that OpenRouter sometimes sends chatbot requests to. The report says an attacker took over a server that was running the DeepSeek V4 Flash 0731 model. Anyone whose requests for that model were routed to Phala between 21 October 2026 at 00:00 UTC and 8 November 2026 at 05:35 UTC should treat the request contents as exposed, and should rotate any API keys, tokens, passwords or other secrets those requests contained.

This is not a new leak from today. The report is about nine days old. The useful part now is checking whether your site's chatbot or AI app was in that window, and rotating anything secret that was in the prompts.

What Phala says happened

Phala says the attacker reached internal management tools on a server on 21 October, and later took control of a new GPU server that was already serving live traffic. That gave the attacker a view of in-flight request data while the model was processing it. Phala says it took that server and the model offline, and that containment was completed on 8 November at 05:35 UTC. It has since restored the service with additional protections. A full root cause analysis has not yet been published.

A third-party researcher had been planting unique test credentials, called canaries, in requests to Phala since 1 July. According to Phala, only canaries from requests sent on 21 and 22 October were later used by someone else. The earliest misuse Phala observed was on 22 October at 05:41 UTC, and the most recent was on 9 November. Phala says those tests sampled requests rather than watching every request, which is why it kept the wider 18-day window.

Why you might have reached that server without choosing it

Privacy Affairs, writing on 11 November, notes that OpenRouter is a gateway that sends a request to one of several companies hosting the same model. Unless you told it otherwise, it picks the provider for you. So you could have reached Phala without ever choosing it.

Privacy Affairs also says this report concerns Phala's own deployment only. DeepSeek's own API, and the other companies serving that model through OpenRouter, are not part of it.

Phala says the gateway is notifying affected users. Privacy Affairs wrote on 11 November that it found no public notice on OpenRouter's own website at the time.

Why this matters if you run a site

Site chatbots and AI-built apps often put API keys, customer details or internal notes inside prompts. A coding assistant can also pull secrets from files you did not mean to paste. You rarely see which company actually handled the request.

We have written before about how quickly exposed keys get used and about what to check when an AI assistant is connected to your tools. The same habit applies here: if a secret was in a prompt, treat it as known to someone else.

What to check today

1. See whether your chatbot or app used OpenRouter in that window

If you have a site chatbot, a coding assistant or an app that sends prompts through OpenRouter, check whether it used the DeepSeek V4 Flash 0731 model between 21 October 2026 at 00:00 UTC and 8 November 2026 at 05:35 UTC. Privacy Affairs says OpenRouter's Activity page shows whether you used the model in that period, and that if your app logs OpenRouter generation IDs, the generation endpoint returns the provider name that served each request.

2. Rotate secrets that were in those prompts

If your requests may have gone to Phala in that window, rotate every API key, token, password or other secret that appeared in the prompts. Include secrets a coding agent may have pulled in from your files, not only what you typed. Then review those accounts for activity you do not recognise, and contact the provider if you find any.

3. Keep secrets out of prompts from now on

The safest secret is the one that never goes into a prompt. Keep API keys, passwords and customer records out of chatbot messages, and keep .env files out of the folders a coding assistant can read.

4. See what a stranger can already reach

A free website security scanner can show what an outsider can see on your site from the internet, without logging in. It will not tell you which prompts were exposed, or which provider handled them. That check still has to be made by you.

FAQ

Did my chatbot definitely leak my passwords?

Only if those passwords, or other secrets, were in the prompts, and only if the requests for that model were routed to Phala in the window above. Phala asks affected users to treat any secrets in those request contents as compromised.

Was this a breach of DeepSeek itself?

No. Privacy Affairs says DeepSeek's own API and the other companies serving this model through OpenRouter are not part of the report. The exposure Phala describes is limited to requests routed to Phala's deployment of that one model.

Why an 18-day window if only two days were confirmed?

Phala says the 21 and 22 October dates come from canary evidence, but the tests sampled requests rather than monitoring every request, and it cannot yet limit the attacker's access to those two days. The wider window is precautionary.

Was this data sold on the dark web?

No source we opened says the prompt data was sold or posted on the dark web. Phala does say that some of the researcher's canary credentials were later misused.

What should a small site owner do now?

Check whether your chatbot or AI app used OpenRouter with that model in the window, rotate any secret that was in those prompts, review the related accounts, and keep secrets out of prompts from now on.

πŸ“° Read the original article on Dev.to Security

Originally published by Dev.to Security. Aggregated on AIWithGhost for educational purposes β€” full credit and traffic to the original publisher.