Cyphrex · 12 checks · About 5 minutes
A checklist for legal, health, finance and insurance teams using AI agents.
Every check comes from a real incident reported in 2026.
One rule: if you can't show proof, answer "No." "Not sure"counts the same way. That's what an auditor, insurer or regulator will do.
3 checks
When an agent uses a person's credentials, everything it does looks like that person did it.
In the news · Anthropic Threat Report
A fraud crew sold "cheap Claude access" whose software harvested customers' logins and resold them. Stolen credentials make an attacker look like the real account holder.
A discount that routes your traffic and credentials through an unknown middleman is a data risk.
In the news · Anthropic Threat Report
Resellers advertising deep discounts on AI access have been powered by stolen accounts and silent model swaps.
Those are two of the warning signs U.S. agencies told companies to watch for.
In the news · Joint Advisory, Sept 8
The NSA, FBI and CISA say six China-based AI companies pulled billions of tokens from U.S. models through pools of accounts and resellers.
4 checks
An overlooked system with broad access becomes the shortcut into everything else.
In the news · GreyNoise via BleepingComputer
AI agents exploited print-management servers at 395 organizations. At one U.S. high school, access became full network control in seven minutes.
An agent with more reach than you planned can act on things you never intended.
In the news · Reuters, Al Jazeera
During a security test with unintended internet access, Google's Gemini got into three real companies it thought were part of the exercise.
A repository's history can hold old passwords, keys and internal notes.
In the news · Reuters
Z.ai's ZCode uploaded entire local code workspaces, full Git history included, to cloud storage by default until users caught it.
The code you point an assistant at can be the attack.
In the news · Accomplish
Researchers found two ways out of OpenAI Codex's sandbox, one of them from read-only mode. Both were patched within eight days.
3 checks
If the most important safeguard is a setting, it's only as strong as whoever can change it.
In the news · Anthropic Threat Report
Anthropic says a warrant requirement in a national surveillance platform was removed at the operator's request, with data set to be kept indefinitely.
Data kept forever is data that can be misused forever.
In the news · Anthropic Threat Report
The same platform's retention was set to indefinite once the control was turned off.
Asking the AI isn't proof. A middleman can fake the model's name in the response.
In the news · Anthropic Threat Report
Customers of the "cheap Claude" reseller were secretly routed to a different model.
2 checks
"The AI did it" won't hold up with a client, a regulator or a court.
Logs you control are a claim. Evidence others can check is proof.
Start answering
Answer the checks above to see where you stand.
0 of 12 verified. 0 of 12 answered.
Cyphrex is an enforcement layer for AI agents. It decides what each agent is allowed to do, records it, and produces evidence anyone can verify independently. Observability tells you what happened. Cyphrex decides what is allowed and proves it.
See how it works at cyphrex.ioFollow @cyphrexio for the stories behind each check