Field guide
Proof of what your AI agent did
10 checks, and a record anyone can verify
Ten checks. Each one opens with something that already happened, then how to check it and how to fix it. The product steps stay in the docs. The point is a record someone else can verify.
01Give the agent its own identity
A fraud crew sold "cheap Claude access" that harvested customers' logins and resold them.
02Buy the key yourself
Resellers advertising deep discounts on AI access were powered by stolen accounts and silent model swaps.
03Catch the account that never sleeps
Six China-based AI companies pulled huge amounts of paid usage from U.S. models through pools of accounts and resellers.
04Cut what it can reach
AI agents exploited print-management servers at 395 organizations. At one U.S. high school, access became full network control in seven minutes.
05Take away the open internet
During a security test with unintended internet access, Google's Gemini got into three real companies it thought were part of the exercise.
06See what the coding tool uploads
Z.ai's ZCode uploaded entire local code workspaces, including every saved version, to cloud storage by default.
07Update it, and cage outside code
Researchers found two ways out of the protected workspace around OpenAI Codex, one while it was supposed to be read-only.
08One person cannot flip the safeguard
A warrant requirement on a national surveillance platform was removed because the operator asked, and the data was set to be kept forever.
09Check which model actually answered
Customers of a "cheap Claude" reseller were secretly routed to a different model.
10Leave a record someone else can check
"The AI did it" will not satisfy a client, an auditor, or a regulator. A log only you can see is a claim.
What it did, whether it was allowed, and a record anyone can check.
Score it in two minutes, or sit down with us and close the gaps. The PDF has the incident, the check, and the fix.
Not legal advice. Incident notes are the ones already on the Cyphrex security checklist. SOC 2 Type I: in progress. Type II: planned. Not certified. Controls are mapped to SOC 2.