HTML and CSS code on a computer monitor, highlighting web development and programming

Time to exploit is now negative seven days: what it means and why AI is shrinking the patch window

By Zain·

Photo: Bibek ghosh / Pexels

CERT-EU says attackers now exploit newly disclosed flaws an estimated seven days before a patch exists, down from 63 days in 2018. Here is how the metric works, what AI changed, and what defenders can still do.

Security teams used to measure how long it took attackers to exploit a flaw after it was disclosed. That number has crossed zero. In an April 21, 2026 post, the EU's computer emergency response team for its institutions, CERT-EU, cited Google's M-Trends 2026 report to say the mean time to exploit newly disclosed vulnerabilities is now an estimated negative seven days, down from 63 days in 2018. The practical meaning, in CERT-EU's words, is that "exploitation typically occurs before a patch exists."

This explainer covers how that metric works, what AI has changed, what the evidence is and is not, and what a defender can do about it. Dates matter here: CERT-EU's post is from April, and the FBI's alert on AI-driven intrusion tools is from November 13, 2026.

What "time to exploit" measures

Time to exploit is the gap between a vulnerability becoming publicly known and the first confirmed exploitation in the wild. Positive numbers mean defenders had a window: a patch could arrive and be applied before attackers moved. A negative number means exploitation was observed before disclosure or before a fix existed, typically because attackers found or bought the flaw first, or because they reverse-engineered a patch within hours.

The estimate is an average and it is not CERT-EU's own measurement. It relies on Google's reporting, and means can be pulled by a few early-exploited flaws. Treat it as direction and order of magnitude: a 63-day cushion in 2018 became no cushion by 2026.

Why patching alone no longer closes the gap

The same CERT-EU post says enterprise patch cycles typically take 30 to 60 days. If exploitation precedes the patch, then even a perfect patching team is exposed for the entire period between disclosure and deployment. CERT-EU adds that exploitation of internet-facing software was the highest-impact initial access vector against Union entities for the second consecutive year, per its Threat Landscape Report 2025, and cites Mandiant's finding that 32% of its 2025 intrusions began with exploits.

What AI changes

CERT-EU argues that AI cuts the cost of both finding and using flaws. It lists several results, each from a different program, so they should not be summed up as one trend line:

“AI-assisted intrusions are highly likely to be faster and stealthier.”

CERT-EU, AI is changing the economics of vulnerability discovery, April 21, 2026
  • Anthropic's Claude Mythos Preview reportedly found thousands of high- and critical-severity vulnerabilities. In a test on Firefox 147's JavaScript engine it produced working shell exploits 181 times out of several hundred attempts, versus two for Claude Opus 4.6.
  • AISLE found 12 of 12 CVEs in the January 2026 OpenSSL release.
  • DARPA's AI Cyber Challenge finalists found 54 vulnerabilities in 54 million lines of code in four hours, at about $152 per task.
  • XBOW submitted nearly 1,060 HackerOne reports, of which 130 were confirmed and resolved.
  • CVE-Bench, a benchmark of end-to-end exploitation, rose from a 13% to a reported 90% success rate.

Read those as capability signals. They are vendor and competition results, and CERT-EU's post is itself a summary. Our tracker of Project Glasswing results shows how many AI-found bugs have been confirmed and fixed, which is the better test of whether the gap runs both ways.

The attacker side, measured in November

By November, the signal moved from benchmarks to incident reports. The FBI's FLASH-20261113-001 says ARTEX, an open-source LLM-driven penetration testing framework, has been used to automate intrusions "at machine speed," with some going from initial targeting to privileged access in under three hours. The Cloud Security Alliance's November 16 note, citing Google's threat intelligence group, gives about four hours on average and 15 minutes in one case. Both describe attackers who do not need to know about a flaw in advance: the agent probes widely, adapts when a path fails and keeps going. We keep a dated log in our ARTEX tracker.

CERT-EU's own prediction, written in April, was that "AI-assisted intrusions are highly likely to be faster and stealthier," and that open-weight models narrowing the gap with frontier systems means these capabilities will likely reach less controlled channels. It also notes that DDoS-for-hire platforms are integrating language models to lower the skill barrier.

What defenders can do

CERT-EU's own approach is instructive: it built an internal AI-powered penetration testing pipeline, reported confirmed findings through coordinated disclosure, and made the pipeline able to generate patches and open merge requests. The general principle is to use the same speed.

The CSA and FBI guidance converges on a short list:

  • Remove default and hardcoded credentials and rotate exposed secrets.
  • Close unauthenticated admin interfaces and inference endpoints.
  • Patch internet-facing software first, since that is the most common entry point.
  • Monitor for bursts of control-plane activity and unusual outbound connections.
  • Revisit incident playbooks, because manual IP blocking may lag agents that change addresses within minutes.

Regulation is catching up slowly. The EU Cyber Resilience Act's vulnerability and incident reporting obligations start in September 2026, with full compliance required by December 2027, per CERT-EU.

What we do not know

No source here gives a measured time-to-exploit figure for AI-found flaws specifically. The negative-seven-days average covers all exploited vulnerabilities and predates much of the November agentic tooling. We also cannot say how much of the shift comes from AI versus faster patch diffing and better attacker economics.

Our take

The metric matters less than the habit it should break. If exploitation can precede the patch, the defender's goal is not to be faster at patching but to have fewer things reachable and fewer secrets lying around when the exploit arrives. The cheapest step for most organizations is boring: inventory what faces the internet, and close what you cannot defend. We would watch the next M-Trends release for whether the average goes further negative, and for the first independent measurement of AI-found flaws.

Frequently asked questions

What does negative time to exploit mean?

It means attackers typically exploit a vulnerability before a patch exists. CERT-EU, citing Google's M-Trends 2026, puts mean time to exploit at an estimated negative seven days, down from 63 days in 2018.

Why can't companies just patch faster?

CERT-EU says enterprise patch cycles typically take 30 to 60 days. If exploitation precedes the patch, even perfect patching leaves a window. Reducing exposure, for example by closing internet-facing services you cannot defend, matters as much.

How is AI changing vulnerability exploitation?

CERT-EU lists AI results that lower the cost of finding and exploiting flaws, from bug-finding competitions to benchmarks, and expects AI-assisted intrusions to be faster and stealthier. The FBI reports some ARTEX intrusions reached privileged access in under three hours.

Is the negative seven days figure reliable?

It is an estimate CERT-EU took from Google's M-Trends 2026 report, not its own measurement. As an average it can be pulled by a few early-exploited flaws. Treat it as direction and scale, not a precise number.

What should defenders do first?

The FBI and Cloud Security Alliance guidance starts with removing default credentials, rotating exposed secrets, closing unauthenticated admin interfaces and patching internet-facing software. They also advise monitoring for bursts of control-plane activity and updating playbooks that rely on manual IP blocking.

Are there rules requiring faster vulnerability reporting?

In the EU, the Cyber Resilience Act's vulnerability and incident reporting obligations start in September 2026, with full compliance required by December 2027, according to CERT-EU's post.

Sources

What each one is, and whose it is.

  1. OtherIndependent of the vendor
  2. 2

    Threat Overview of ARTEX Autonomous Penetration Testing Framework (FLASH-20261113-001), FBI Internet Crime Complaint Center (November 13, 2026)

    OtherIndependent of the vendor
  3. 3

    ARTEX and Its Forks: Agentic Pentest Frameworks as Attack Infrastructure, Cloud Security Alliance AI Safety Initiative (November 16, 2026)

    OtherIndependent of the vendor