Earlier this year an AI research agent found a critical bug in BeyondTrust's remote access products. Four days after it was disclosed, attackers were already using it. Within a week, six separate groups were.
That number comes from Google's Threat Intelligence Group, which put out a report on September 30 called Vulnerability Discovery and Exploitation Trends in the AI Era. It's worth reading the whole thing. Here's the short version, and what I think it means if you run a small web app.
What happened
The bug is CVE-2026-1731, an unauthenticated OS command injection in BeyondTrust Privileged Remote Access and Remote Support. No login needed. Send the right request and you're running commands on the box.
Google says it was found autonomously by Hacktron AI, a third-party research agent. That part is good news. A machine found a serious bug and it got reported and fixed. The bad news is what came next: according to GTIG, one threat cluster was exploiting it within four days of public disclosure, and five more joined within seven days.
The bugs AI finds are the ones attackers want
This is the line from the report that stuck with me. Google found that exactly 50% of AI-discovered vulnerabilities lead to remote code execution. Across all CVEs, it's 26%.
So AI isn't just finding more bugs. It's finding the worst kind, roughly twice as often. Those are exactly the ones attackers race to use.
The rest of the numbers point the same way:
- Monthly CVE disclosures doubled this year, from 5,045 in January to 10,740 in August.
- GTIG recorded 141 vulnerabilities disclosed and exploited from January to August 2026. That's already more than all of 2025, which had 127.
- Only 0.23% of 2026 disclosures, about 1 in 431, were seen exploited in the wild.
That last one matters. Most CVEs never get used. The few that do get used fast.
Patch everything is not a plan
When disclosures double, nobody keeps up by patching harder. Google says it plainly: teams need to "transition from unprioritized mass-patching to threat-intelligence-driven triage."
For a big company that means a threat intel team. For a five-person startup it means something simpler. You need to know which of your things face the internet, and you need to move fast on those when a critical bug drops. Everything else can wait for the normal cycle.
What I'd do this week
- Write down everything that's reachable from the internet. Your app, admin panels, VPN, remote access tools, the staging server someone spun up in March. If you don't know it's there, you won't patch it in four days.
- Watch CISA's Known Exploited Vulnerabilities list. The KEV catalog is free and short. If something on your list shows up there, that is your drop-everything signal.
- Turn on automatic updates where you can. Managed hosting, OS security patches, dependency bots. Four days is less than most teams take to review a pull request.
- Pay extra attention to AI tooling. Google tracked 212 vulnerabilities disclosed this year in inference and serving software like vLLM, Ollama, and LiteLLM, and nearly a quarter were unauthenticated API endpoints or SSRF. If you're self-hosting any of it, it belongs at the top of your list.
- Shrink what's exposed. Admin tools and remote access products don't need to be open to the whole internet. Put them behind a VPN or an identity-aware proxy and a four-day exploit window stops being your problem.
The bottom line
Attackers have the same AI tools defenders do, and they're turning new bugs into working attacks in days. The window where "we'll get to it next sprint" was safe is closing.
If you're not sure what you have exposed, start with the free scan. It takes a couple of minutes and shows you what the internet can see. A full pen test goes further and tells you which of those things an attacker could actually use.