Researchers told three AI inference providers this week that attackers had been stealing customer prompts from inside their hosts for months. None of the three knew. OpenRouter, the gateway that routes customer model requests into a pool of those hosts, dropped the providers after the industrial cybersecurity company Dragos made the theft public. Some of the stolen requests were flagged Zero Data Retention, the setting that is supposed to mean a prompt is never stored.
You already paid for that flag. Legal already treated it as a control. The host kept the text, and your tooling never saw the copy because the copy never crossed a network you operate.

Prompt Theft Persisted for Months Under Zero-Retention Flags
OpenRouter’s job is routing. Customers hit one API. The request lands on an inference host they didn’t select by name, under a wrapper that looks like a single vendor. That design is why people use a gateway. It’s also why a breach at a provider you never contracted with can still hold your prompts.
Zero Data Retention is a policy bit on a request. It’s not encryption you hold the keys for. If a worker process, a debug log, a cache, or a stolen credential on the host retains the payload, the bit did nothing. Dragos researchers found months of theft. The providers’ own threat detection didn’t surface it. You learned because someone else knocked.
Prompt text is a gift once an attacker has a foothold. Engineers paste stack traces, IAM drafts, database URLs, and the production token they meant to rotate. Support pastes customer tickets. A prompt archive is recon and credential material in one stolen table.
From your side the session looked clean. The firewall allowed TLS to a known AI hostname. Threat-protection tools scored it as sanctioned SaaS. Brute-force dashboards stayed quiet because nobody was guessing passwords on your edge. The store that mattered sat on a host you can’t image, under a retention promise you can’t verify.
This is a bad look for every inference vendor that sells “we never store your prompts” as a dashboard toggle. The real problem here is operational. You can’t attest to a control that lives entirely on someone else’s GPU box unless you can test it.
Crafted Uploads Escaped Ghostscript Sandbox to Host RCE
The same week gave you the file-conversion version of that failure. Security firm pwn.ai showed a now-patched Ghostscript flaw that lets a crafted file escape the interpreter’s sandbox and run commands on the server that processed it. Ghostscript is how a lot of Linux servers turn uploaded PDFs, PostScript, and images into other formats. Popular apps often invoke it without naming it in application code.
Ordinary upload features are the delivery path. Attach a PDF. Render a CMS thumbnail. Preview the first page of a ticket attachment. The HTTP handler checks size. A helper then hands the bytes to Ghostscript. If the sandbox doesn’t hold, that helper is an RCE gadget with a document as the exploit.

Treat document converters as interpreters, because they are. A sandbox around PostScript is still a language runtime on a host with network access, cloud credentials, and a view of every other upload in the spool. When that host shares a kernel with the web app, a thumbnail request becomes a shell.
Operators Must Inventory Indirect Processors This Week
If a gateway, a model host, or a file converter can see the payload, it belongs in your cyber security scope the same way an identity provider does. Feature teams will call these integrations. Put them on the asset list as processors, and give them owners.
Do the inventory this week, then repeat it when someone adds another model to a chatbot.
- List every AI gateway, model endpoint, and third-party inference host your apps call, including shadow tools on developer laptops.
- Record whether Zero Data Retention is a contract clause, a request header, both, or neither, and who can prove it.
- Plant canary strings in test prompts and watch DNS, email, and secret-scanning for callbacks.
- Run Ghostscript, ImageMagick, Poppler, and LibreOffice under a dedicated user, with no cloud credentials, on a host that can’t reach production data stores; disable internet-facing preview until that isolation exists.
Rotate any credential, customer identifier, or architecture detail you know went into prompts on the dropped providers. Treat gateway logs as incomplete. Ask remaining hosts for processor lists, subprocessors, retention timestamps, and whether ZDR traffic was excluded from the stolen set. Get it in writing. If they can’t say, assume the prompts were kept.
Ongoing security hardening is egress control, not a new banner in the chatbot UI. Pin model endpoints. Block unknown inference hostnames from production and from the laptop VPN. Keep conversion workers on a dedicated subnet with no inbound SSH from the internet and no role that can read the main object store. Give those workers a patch SLA in hours for interpreter CVEs. The next sandbox escape will arrive as a file.
Incident response needs a trigger that isn’t a reporter. Pre-negotiate notification windows, forensic artifacts, and the right to pull your data. Tabletop the version where the vendor has no logs because ZDR forbade logging, and the attacker has the prompts anyway.
Cybersecurity Monitoring Did Not Reach Inference Hosts
Defense in depth on the slide deck still ends at your edge. You instrument the firewall, the IdP, the EDR agent, and the SIEM parse for brute-force against SSH and VPN. That work is real. It had no sensor inside the inference worker that stored a never-store prompt, and no sensor inside Ghostscript’s sandbox. Those were other people’s processes, holding your data, on someone else’s patch cycle.
Log the destination host or provider ID your router actually selected, not only the front-door hostname. Alert when production identities call a model host that isn’t on the allow list. Alert when a conversion worker makes an outbound connection; a thumbnail process has no business talking to the internet. Watch for worker crashes and sudden preview latency. That’s how malformed-file probing often shows up before a CVE lands.
Threat detection for this class of failure is slow if you wait for the vendor. Assume prompts and uploads are already copies the moment they leave your process. Classify them that way. Limit what staff may paste. Strip secrets in the client before the request hits the gateway. If the business insists on sending source to a pooled model, send it to a host you run, on hardware you can wipe.
You’re going to keep using AI gateways and PDF previews. Stop treating a routing product and a converter sandbox as a security boundary. The boundary is the last system you can image, the last log you can pull, and the last retention claim you can test. Everything past that is someone else’s incident, running on your data, on a timeline you don’t set.
Sources
- OpenRouter Drops 3 Providers After Researchers Catch Months of Prompt Theft at AI Hosting Firms
- Researchers Demonstrate Ghostscript Sandbox-Escape RCE via Crafted File Uploads in Popular Apps
Take Control of Your Server Security
Don't let brute-force attacks slow you down. Try IPBan Pro risk-free for 30 days.
Secure. Automated. Lightweight.
Take Control of Your Server Security
Don't let brute-force attacks slow you down. Try IPBan Pro risk-free for 30 days.
Secure. Automated. Lightweight.
