Part 3: Predicting the exploit
Part 3 of 6 in a deep dive on machine learning that predicts attacks and stops them early, with Apple as the main case. Start with Part 1 if you are new to the series.
The patch window closed
Mandiant measures the mean time between a vulnerability's public disclosure and the first time it is seen exploited. It was 32 days in 2021 and 2022 and 5 days in 2023 (TechRepublic on Mandiant). In 2024 it fell to minus one day (Tenable), and for 2025 Mandiant estimates minus seven days, which means exploitation now routinely starts before a patch exists (Google Cloud).
The rest of the data points the same way. Of the 138 exploited vulnerabilities Mandiant analyzed for 2023, 70% were zero-days, used before any patch existed (TechRepublic). Google counted 75 zero-days exploited in the wild in 2024, 44% of them in enterprise products such as VPNs and firewalls (DeepStrike on Google data). In 2025 Mandiant saw the handoff from the group that breaks in to the ransomware crew that cashes out fall from 8 hours to 22 seconds (Help Net Security).
A patch cycle measured in weeks cannot keep up with a negative number.
2026: attackers got AI
The numbers above mostly describe 2025. The reports published in 2026 show why the curve keeps bending. Attackers now use AI models at almost every step, from writing the lure to running the intrusion.
AI inside the attack
Google's Threat Intelligence Group publishes regular updates on how attackers use AI. In February 2026 it described malware families such as HONESTCUE that call Gemini while they run and ask for new code, which moves the malicious logic out of the file a scanner sees, and a rise in attempts to copy commercial models by querying them at scale (SiliconANGLE). In May it reported the first case it had seen of a criminal group using a zero-day exploit that GTIG believes was developed with AI. Google blocked the attack before it succeeded (KnowBe4).
In its tracker for the third quarter, GTIG's chief analyst John Hultquist said defenders can now assume that every threat actor uses AI in some way. Mandiant saw one attacker compromise an organization's cloud, deploy an autonomous multi-agent attack framework there and harvest thousands of credentials within six hours. Another group, tracked as TeamPCP, backdoored MCP tools used by AI coding assistants and poisoned package metadata so that assistants would recommend malicious dependencies (Cyber Magazine).
Social attacks scale up
The biggest change is in social engineering, because AI removes the effort that used to limit it. GTIG saw Russia's Sandworm use Gemini for social engineering and password-spraying scripts against Ukraine, and Iranian groups use it to design photorealistic fake personas (Cyber Magazine). CrowdStrike's 2026 report counted an 89% rise in attacks by AI-enabled adversaries, a 563% rise in incidents using fake CAPTCHA lures and 141% more spam (CrowdStrike). Mandiant found that voice phishing was the most common way into cloud environments, at 23% of the cloud compromises it investigated (Google Cloud community).
The FBI's Internet Crime Report, published in April 2026, broke out AI for the first time in its 25 years: 22,364 complaints and $893 million in losses, out of a record of more than one million complaints and nearly $21 billion in total. Scammers used AI for fake social profiles, cloned voices, forged documents and videos of public figures or relatives (FOX, FBI release). In Gartner's 2026 survey of CISOs, 41% reported a deepfake combined with social engineering on an audio call, and 35% on a video call (Keepnet on Gartner).
Less time once they are in
CrowdStrike measures breakout time, the gap between an attacker's first foothold and their first move to another system. It has fallen every year, and in one 2025 intrusion data exfiltration started four minutes after initial access (CrowdStrike press release).
The same models on defense
Defenders have the same tools. Anthropic launched Project Glasswing on April 7, 2026, giving partners including AWS, Apple, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA and Palo Alto Networks access to its Claude Mythos Preview model for defensive vulnerability research. In its first month the project flagged more than 10,000 high- or critical-severity vulnerabilities in over 1,000 open-source projects (Daily Security Review). I covered the start of Glasswing in my second deep dive.
The bottleneck moved to fixing. Of roughly 1,596 vulnerabilities Anthropic had disclosed to open-source maintainers by May, 97 were patched, about 6% (Cloud Security Alliance). The Zero Day Clock project, which tracks the gap between disclosure and exploitation, puts it at roughly 20 hours in 2026 (TechRadar). Finding bugs is no longer the scarce part. Deciding which ones to fix first is, and that is where prediction comes in.
Predicting which bug gets exploited
In 2025 more than 48,000 vulnerabilities received CVE numbers, and 884 were added to CISA's catalog of known exploited vulnerabilities (DefenderMate). Nobody can patch 48,000 things at once, so the question is which ones first. The common answer has been severity: patch everything with a CVSS score of 7 or higher.
The Exploit Prediction Scoring System, EPSS, answers the question with machine learning. Its authors presented the first model at Black Hat in 2019 (arXiv), and FIRST now publishes a daily score for every published vulnerability: the probability that it will be exploited in the next 30 days (SecPod). Version 4 came out on March 17, 2025. It learns from evidence of around 12,000 vulnerabilities exploited every month (Empirical Security).
Jay Jacobs, who has worked on EPSS since the start, compared the two approaches. Patching everything at CVSS 7 or above means working on 50.7% of all published vulnerabilities, and only 6% of that effort lands on vulnerabilities that are actually exploited. Holding coverage roughly the same, EPSS v4 needs 6% of the effort, and 47% of it lands on exploited vulnerabilities (Empirical Security).
EPSS is the same move as scoring a sign-in in Part 4. A model turns a pile of equal-looking items into a short list. It knows nothing about your network, so it works best combined with what you do know: whether a system faces the internet, what data it holds, and whether a bug is already on CISA's known-exploited list (SecPod).
Finding the bug before the attacker
Prediction can go one step earlier, to bugs nobody has published yet.
Big Sleep finds its first real bug
Google's Big Sleep agent, built by Project Zero and DeepMind, finds a stack buffer underflow in SQLite that could have led to a crash or code execution (The Hacker News).
Big Sleep stops an exploit in the wild
Google's threat intelligence group sees attackers staging a zero-day but cannot tell which bug. Big Sleep finds it, CVE-2025-6965 in SQLite, and it is fixed in version 3.50.2 before it is used. Google calls it the first time an AI agent directly foiled exploitation in the wild (The Hacker News).
CodeMender writes the fixes
Google DeepMind's CodeMender agent upstreams 72 security fixes to open-source projects in six months, in codebases of up to 4.5 million lines, with every patch reviewed by people (The Hacker News). It also adds bounds-safety annotations to libwebp, the image library linked to a 2023 zero-click iPhone exploit (Ground News).
Apple attacks its own design
Apple's offensive research team spends five years attacking Memory Integrity Enforcement before it ships, first on paper and then on prototype chips (Apple Security Research).
The pattern is the same in each case. The defender looks for the attack before the attacker finishes building it, whether the tool is a probability score, an AI agent or a team paid to break the design. Prediction tells you which door attackers will try next. It does not stop the attacker who is already pulling on the handle. Part 4 turns to the locks themselves, the models that decide in seconds whether a sign-in, a request or a file gets through, and what happens when attackers go after those locks.
Continue to Part 4: Locks at machine speed