Ivanti Used AI to Find a Perfect-Score Security Flaw in Its Own Software. Here Is What That Means.
The company quietly ran an AI project starting in March and says the results are already surprising even its own security chief.
Key points
- Ivanti disclosed CVE-2026-10520, a critical flaw in its Sentry mobile gateway product, with a perfect severity score of 10 out of 10.
- An AI language model, not a human researcher or engineer, discovered the flaw.
- Ivanti began its internal AI security project in the first week of March 2025, using models from Anthropic and OpenAI.
- Ivanti's chief security officer told Dark Reading he suspects a recent bug-bounty submission was written by an AI, potentially funded by a roughly $100-per-month software subscription.
- The company plans to release research figures on vulnerabilities found and fixed in the coming months.
A software company used an AI language model to find a critical flaw in its own product before any criminal did. That company is Ivanti, which makes networking and security software used by businesses and governments. We first covered CVE-2026-10520 on 10 June 2026, and our earlier reporting on the same Ivanti patch cycle shows this vulnerability did not arrive alone.
The flaw sits inside Ivanti's Sentry product, a gateway that controls access from mobile devices to corporate systems. It received a CVSS, or Common Vulnerability Scoring System, score of 10 out of 10, the highest possible rating. Attackers who found it first could have taken full control of affected systems. No human found it.
How did Ivanti get an AI to hunt for security flaws?
Ivanti's chief security officer, Daniel Spicer, says the project started informally in mid- to late February 2025, when he and a colleague noticed that the 4.6 generation of Anthropic's Claude models had become capable enough to handle real security work. The formal project launched in the first week of March. It ran on two tracks: one to find flaws that existing automated scanners were missing, another to fix flaws automatically before a human engineer had to touch them.
On the fixing side, the process goes like this. Ivanti's static and dynamic code-scanning tools flag a weakness. Instead of routing it to an engineer, an AI agent picks up the problem, writes corrected code, and passes it back to the scanner to confirm the flaw is gone. If the scanner agrees, automated tests check the fix hasn't broken anything else. Only then does the code return to an engineer.
Spicer told Dark Reading the AI operates inside a tightly limited space: it receives a single clearly defined task and can't access Ivanti's wider systems. That design is the right call. Our story "AI Finds Bugs Fast. Proving They're Real Still Takes a Human." makes the case that unconstrained agents are where defensive AI projects tend to come apart.
Ivanti is using models from Anthropic and OpenAI, plus some open-source models run internally.
Should you worry about AI on the attacker's side too?
The same capability cuts both ways. Spicer said a bug-bounty submission arrived recently, where the formatting looked like output from an Anthropic Opus model. The flaw was one Ivanti had already spotted, so no harm was done. But the economics are stark: a $100-per-month AI subscription could help someone earn a $5,000 to $10,000 bug-bounty payout.
That's the part of this story worth watching most closely. Ivanti using AI to find its own critical flaws is genuinely good news. But Spicer's bug-bounty anecdote is a signal that the same tools are already lowering the barrier for opportunistic researchers, and possibly for less honest actors, to do the same work at scale.
If your organisation uses Ivanti Sentry, confirm with your IT team that the patch covering CVE-2026-10520 has been applied. If a corporate mobile device management system is part of your setup, ask whether Sentry is in the chain and whether it's current.
