For twenty years, security teams have used one quiet shortcut to decide what to patch first: is there a public exploit? If not, the bug could usually wait. Anyday Security, a New York start-up with an R&D center in Tel Aviv, launched a free site on November 11 2026 that is built to break that habit.
ExploitTotal.ai is a catalog of working exploits written by an open-weight AI model. At launch it holds more than 200. Anyday says about 72% of them target vulnerabilities that never had a public exploit before. It checked each one against the US National Vulnerability Database, CISA’s Known Exploited Vulnerabilities catalog and public exploit repositories.
The method is the point. Anyday gave the model only what any attacker gets on Patch Tuesday: the vendor’s public vulnerability description and the patch itself. The model it used is GLM-5.3, which the company says the US Center for AI Standards and Innovation called the most cyber-capable open-weight model to date. Anyone can download and run it. In one demonstration, Anyday says, the model built a full-chain browser exploit in minutes for less than a dollar.
What defenders actually get
Each vulnerability page gives a verdict on whether AI models can weaponize the bug, the cost and time it took, the model that did best, the severity, and whether a public exploit exists. Results are graded on a five-level scale, from crashing the target program to full control of it, and tagged with model and date. That last detail turns the catalog into a running record of how fast AI exploit capability is improving.
Access is tiered. Anyone can search. Signed-in users see the full exploitability assessment. Only verified users can download proofs of concept, and those are unweaponized: they trigger the bug against vulnerable code without a harmful payload. Anyday says they cover patched, publicly disclosed CVEs and are meant for authorized testing, research and building detection rules. Coverage starts with the software enterprises run most, including Windows, Linux, Chrome, Firefox and Edge.
Why the patch window just shrank
The numbers in Anyday’s release explain the urgency. Microsoft fixed 1,411 vulnerabilities in its October 2026 Patch Tuesday, more than in all of 2025. Verizon’s 2026 Data Breach Investigations Report puts the median time to fix actively exploited flaws at 43 days, with only 26% fully fixed. Google’s threat intelligence group counts an average of 18 exploits a month in 2026, nearly double last year.
Put those together and the old triage rule fails. If a free model can turn a patch into a working exploit in minutes, then “no public exploit” no longer means “no exploit.” It means nobody has posted one yet. Co-founder and CEO Yonatan Gorfung says the same, more bluntly.
The founders are both veterans of the Israeli military’s Unit 8200. CTO Omri Sibony was previously director of security research at SentinelOne. The release does not mention funding.
The obvious objection
A free catalog showing which bugs AI can weaponize is useful to attackers too. Anyday’s answer is the access model: verdicts are public, downloadable code is gated and payload-free, and everything covers bugs that already have patches. Attackers with GLM-5.3 can already do this themselves. The catalog mostly hands defenders the same knowledge.
For security teams, the practical change is in the patch queue. Stop ranking by “exploit exists in the wild.” Rank by “can a cheap model build one,” and assume the answer for browsers and operating systems is now yes within hours of the patch.
For the ransomware numbers behind this pressure, see opint.com.