# An LLM Found a CVSS 10 Bug Before Any Human Did — Ivanti's AI Bet Is Paying Off
## The Threat
When Ivanti disclosed CVE-2026-10520 last month — a critical, maximum-severity flaw in its Sentry mobile gateway product — the announcement contained an unusual detail buried in the fine print: the vulnerability wasn't found by a researcher, a bug bounty hunter, or even Ivanti's own engineers. A large language model flagged it first.
That fact alone reframes the conversation around AI and security. The dominant narrative of 2025 and 2026 has been threat actors weaponizing AI to write better phishing lures and faster malware. The counternarrative — defenders using the same frontier models to hunt and patch their own code — has moved more quietly. Ivanti's disclosure changes that.
The vulnerability affects Sentry, Ivanti's mobile gateway used by enterprises to proxy and secure traffic between mobile devices and backend systems. A CVSS score of 10 — the ceiling — indicates remote exploitability, no authentication required, and full system compromise potential. The fact that traditional tooling missed it, and that an LLM caught it during an internal red team exercise, raises uncomfortable questions about how many similar flaws are sitting undiscovered in vendor code right now.
## Severity and Impact
| CVE | CVSS Score | Product | Attack Complexity | Authentication | Privileges Required |
|-----|-----------|---------|------------------|----------------|---------------------|
| CVE-2026-10520 | 10.0 (Critical) | Ivanti Sentry | Low | None | None |
*Note: Full CVSS vector string and CWE classification were not released in Ivanti's initial disclosure. This table reflects confirmed scoring at time of publication.*
## Affected Products
- Used by enterprises for mobile device management and backend traffic proxying
- Specific affected version ranges to be confirmed in vendor advisory
*Organizations running Ivanti Sentry in any capacity should treat this as urgent. Given the CVSS 10 rating, assume exploitation is feasible by any motivated attacker with network access.*
## Mitigations
Patch immediately. Ivanti released a fix alongside the June 2026 disclosure. There is no responsible case for deferring a CVSS 10 patch on a perimeter-facing gateway product.
While patching is underway:
Ivanti has provided remediation guidance through its customer portal. If your Sentry deployment is managed through a third-party MDM vendor, confirm with them that the patched version has been deployed in your environment.
## References
---
## HackWire Analysis
The story here isn't just one critical CVE — it's what finding that CVE says about where vulnerability research is heading.
Ivanti CSO Daniel Spicer told Dark Reading that the program kicked off in early March 2026, roughly a month after he and a colleague noticed that Claude's 4.6-generation models had crossed a threshold of usefulness for security work. Two tracks were stood up simultaneously: one aimed at automatically resolving known vulnerabilities, the other at surfacing flaws that standard tooling was missing. CVE-2026-10520 came out of the second track.
That's the detail worth sitting with. Traditional static analysis, fuzzing, and human red-teaming — the standard toolkit — all missed a maximum-severity bug in a perimeter product that enterprises rely on for mobile security. An LLM running as part of an augmented red team caught it. That's not a fluke to be explained away; it's a signal about the nature of the vulnerability classes frontier models are good at finding. Likely: logic flaws, authentication bypasses, and subtle protocol-level issues that don't pattern-match to signatures but do look wrong to a model trained on millions of lines of secure and insecure code simultaneously.
The real question for the industry is economic. Ivanti can absorb the inference costs to run this program because the liability exposure from unpatched critical CVEs is enormous. Smaller vendors can't — and their products sit in the same enterprise environments. We should expect pressure to build for shared vulnerability research infrastructure, or for major cloud providers to offer LLM-assisted security review as a bundled service. The alternative is a widening gap between well-resourced vendors patching their code with AI assistance, and smaller players who ship the flaws their tools can't see.
For defenders right now: patch CVE-2026-10520, obviously. But also watch Ivanti's disclosure cadence over the next two quarters. If they're running this program at scale, more AI-found CVEs are coming — and the severity distribution may be surprising.
— HackWire Editorial
## Related Coverage