Dateline: July 21, 2026
An AI Found the Flaw First. Then Attackers Found It Too.
A software vendor pointed a large language model at its own code. The model found a hole so bad it scored a perfect 10 out of 10. No human researcher spotted it. No scanner flagged it. A machine did. Then, within 24 hours of the fix going public, attackers were already breaking in.
What Happened?
Ivanti disclosed CVE-2026-10520 on June 9, 2026. It is a command injection flaw in Ivanti Sentry, the gateway that sits between mobile devices and a company’s internal email and apps. The flaw let a remote attacker run code as root without logging in first. That is about as bad as it gets.
Here is the part that turned heads. Ivanti did not find this bug the usual way. Chief Security Officer Daniel Spicer told Dark Reading that his team started using frontier AI models in March to hunt for and fix flaws. This one slipped past their normal tools. The model caught it.
Ivanti runs two tracks. One track uses AI to find bugs that scanners miss. The other uses AI to fix flagged weaknesses before an engineer even sees the ticket. Spicer said he was pleasantly surprised by the fix quality, and that both Anthropic and OpenAI models pulled their weight.
But the win came with a catch. Researchers at watchTowr published a technical breakdown and a proof of concept on June 10. That handed attackers a map. On June 11, CISA added the flaw to its Known Exploited Vulnerabilities list. One report found at least two vulnerable systems already backdoored.
What’s the Impact?
Ivanti Sentry is often exposed to the internet. Crack it open and an attacker can grab credentials, steal session tokens, and pretend to be a real user. From there they reach email servers and internal apps. Ivanti also patched a second flaw, CVE-2026-10523, an authentication bypass rated 9.9 that lets attackers create their own admin accounts.
Spicer made a sharper point too. The same models that help defenders also help attackers. He said his team received a bug bounty report that looked machine written, right down to the formatting. A $100 monthly AI subscription, he noted, could net someone a $5,000 to $10,000 bounty. The economics of finding flaws just shifted for everyone, good and bad.
How to Avoid This
Patch now, outside your normal schedule. Upgrade Sentry to R10.5.2, R10.6.2, or R10.7.1.
- Check port 8443. The management interface should never face the open internet.
- Review admin accounts. Look for any created without a clear reason.
- The bypass flaw exists to make fake ones.
- Run watchTowr’s public detection script to check if you are exposed.
- Assume compromise if you were exposed before patching.
- Rotate credentials and session tokens tied to that gateway.
When the Scanner Can’t See It, Something Else Should
The Ivanti case shows a gap most teams already know about. Traditional scanners follow the code and miss the context. They cannot tell that an endpoint should have required a login. That is where a Digital Security Teammate fits.
Secure.com helps close that gap without adding another blind tool to the pile.
- Its AppSec Teammate reasons about code the way a reviewer would, catching missing authorization checks that static scanners walk right past.
- Its Infrastructure Security Teammate keeps a live view of exposed assets, so an internet facing gateway does not sit forgotten.
- It ties findings to real risk, so your team patches the 10 out of 10 flaw first, not the noise.
- It shortens the window between a flaw going public and your fix landing, which is exactly the window attackers raced through here.
