What is AI actually doing to real-world security?
This site tracks the outcomes underneath the arguments about AI: breaches and what they cost, how fast disclosed bugs get exploited, and the documented cases of AI in attack and defense, with a decade of vulnerability data beneath it. Evidence of AI's involvement is graded by how well it is confirmed, and where no source can see AI's role yet, the page says so.
Three kinds of AI: closed models, open-weight models and attack harnesses. The site sets out the true state of each, as far as the evidence reaches.
Losses reported to the FBI rose 16× from $1.3B in 2016 to $20.9B in 2025; average eCrime breakout fell from 98 to 29 minutes
Two outcome measures on one year axis, each on its own scale and from its own source: money lost in complaints filed with the FBI's IC3, and how fast attackers in CrowdStrike's eCrime cases moved from first access to the rest of the network. IC3 tags $893M+ of its 2025 losses as AI-related, a lower bound that counts only victims who noticed AI and said so; the rest of its total carries no AI field. CrowdStrike asserts that AI accelerated phishing and automated reconnaissance, shortening the time from initial access to impact.
(2026 Global Threat Report, PDF p.2). That is CrowdStrike's own claim, with no measure of how much of the drop AI caused.
Method and caveats
Several IC3 editions state their total as a rounded floor ("exceeding $3.5 billion"); the bars use the stated figure. CrowdStrike's earlier editions measured other populations, so only its eCrime figures are drawn. IC3's AI-tagged losses, a lower bound counting only victims who noticed AI and said so, are on Outcomes.CVEs per year rose 2.9× from 2018 to 2025; average time to exploit fell from 63 days to 5 by 2023, and M-Trends 2026 puts it at −1 day in 2024 and an estimated −7 days in 2025
Two measures on one year axis, each on its own scale and from its own source. They show timing, not cause: none of these series attributes any change to AI. The hollow spans come from a different publication with a different definition: M-Trends 2026 reports Google Threat Intelligence Group's mean time to exploit, where a value below zero means exploitation came before the patch was released, over a set of vulnerabilities the edition does not state, and it labels 2025 an estimate. The filled spans are averages from Mandiant's 2023 analysis of the vulnerabilities it saw exploited; M-Trends 2026 gives 63 days for 2018, the figure Mandiant gave for 2018–19. The two are drawn side by side, never joined into one series.
Method and caveats
How Mandiant computed them: The AI Question. The 2017 jump and the 2024 surge in CVEs trace mostly to more issuers, not more bugs: see CVE Data. A public, monthly time-to-exploit measure built from CISA KEV is on Live; it counts to CISA's listing, so it runs longer than Mandiant's figures.Short version
Breach notices, ransomware leak-site postings and the speed at which disclosed bugs get exploited are measured here from public sources. None of those sources records whether AI was involved, so this site attributes none of their movement to AI. What is documented: AI systems now find real bugs, and the ledger holds 23 documented AI-assisted attack campaigns: 6 confirmed by someone other than the model provider, 13 resting only on the model provider's own report and 4 asserted without evidence a reader can check. How much AI has moved breaches, losses and disruption is the question this site measures, one graded finding at a time. See the full reasoning →
The attack count reads 14 published documents: 7 model providers' reports (OpenAI, Anthropic and Google GTIG), a victim's own statement (Arup), 2 security researchers' reports (HP Wolf Security and Zenity Labs), a CERT advisory (CERT-UA) and 3 incident-response and threat-intel reports (Wiz, Palo Alto Networks Unit 42 and Cisco Talos). It holds what those documents describe, so it is a partial record, not a census of AI-assisted attacks.
The one-paragraph answer
Start with outcomes. Breach notices, ransomware leak-site victims and the time from disclosure to confirmed exploitation are tracked here month by month, and exploitation became the most common breach entry vector in Verizon's 2026 DBIR. None of those series says whether AI was involved, so none of their movement is attributed to AI here; that is a limit of the instruments, and the site says so wherever it applies. Underneath sits the vulnerability supply. CVE volume is climbing steeply. Documented contributors that predate AI are part of it: more CVE issuers, high-volume trackers such as the Linux kernel and WordPress plugins, and more complete disclosure. No public breakdown says how much of the rise each one explains, AI included. How the mix of vulnerability types is shifting is computed on Analysis. What is documented about AI itself: systems such as Big Sleep, AISLE and DARPA's AIxCC finalists find real bugs, and the ledger holds 23 documented AI-assisted attack campaigns: 6 confirmed by someone other than the model provider, 13 resting only on the model provider's own report and 4 asserted without evidence a reader can check. The strongest attacker-side claims (autonomous AI-run campaigns) remain contested. Interpretation: the best-documented cases so far are defensive uses, such as AI finding vulnerabilities, and the outcome evidence cannot yet say which way AI has moved real-world harm. Flat breach counts do not settle it either: they include only breaches someone found, and in October 2026 incident-response firms reported turning work away (Defender capacity).
What's on this site
How to read this. Every number is tied to a named source: per-figure links live in the data files and on the Sources page. Where sources disagree, both are shown. Where the data can't answer a question, it says so instead of guessing. Vendor marketing is labeled as vendor marketing. Neutrality is the point: AI's effect on an outcome is stated only as far as the evidence behind it allows, and no page declares a net winner.