Drift ·
The half of a pentest nobody keeps
A pentest finds the vulnerability and hands you a PDF. This release is about the other half — turning each finding into the thing that keeps watching for it.
Run a scan or a pentest and you get a list of vulnerabilities and a date. Both decay. The code changes, the agent gets a new tool, someone tries the same trick again next week — and the finding, which was true the day it was written, quietly stops being the thing you need. The expensive, boring, valuable half of security isn't finding a vulnerability once. It's keeping it found.
Why most tools stop at the finding
A pentest tool attacks from the outside and leaves. It was never on the machine, so it can't watch anything after it goes. The autonomous pentesters are good at the first half — they'll tell you where the hole is. They structurally can't do the second, because they aren't running when the next attempt arrives. And for AI agents specifically, the most dangerous vulnerabilities aren't reachable from the outside at all: a malicious instruction hidden in a document the agent reads, a tool whose description quietly tells it to exfiltrate, an agent that takes a high-impact action because some text told it to. You can only test those by driving the real agent.
What changed here
AariaSec has always run continuous behavioral monitoring on the machine — it learns how each agent normally behaves and flags the one that stops behaving like itself. This release adds the offensive half and wires the two together. It drives your real AI agents, tests the vulnerabilities an external scanner can't reach, and proves which ones actually work. Then — the part that matters — it hands you the runtime rule already watching for that exact behavior. For the agent-level traps, it can tell you whether the monitor already caught the exploit when it happened.
That's the loop a scanner can't close, because a scanner was never going to be there tomorrow.
The finding is free. Watching it is the point.
You shouldn't have to pay to be told you have a problem, so the whole offensive loop — scan, proof, real-world impact, the fix, and the proposed runtime rule — is free. You pay when you want it enforced and governed: automatic containment, cost limits, four-eyes approval before an agent switches model, per-task kernel scope-lock. Free finds and watches; the paid tier contains and governs.
On the way, the scan flags the non-AI exposures it meets too — leaked credentials, weak DNS/TLS/DMARC, risky config. We're not claiming to be your whole security program. We disclose what we find.
The same question, extended
It's the question we've always asked — does what an agent does match what it's supposed to do — now reaching from watching into hunting. The finding was always the easy half. Run a free Security Check and keep the half that matters.