
Frontier AI models are shrinking the time attackers need to exploit software flaws to hours, but organisations are still taking weeks or months to fix them, according to new research from Black Duck.
The company’s 2027 AI Readiness Report, published today, analysed 65.5 million application security findings across around 13,100 applications at 533 organisations between June 2025 and May 2026. It found that organisations were discovering 18 new vulnerabilities for every one they fixed by May 2026, up from 11 to 1 a year earlier. The ratio peaked at 22 to 1 in December 2025.
Only 10.5% of the vulnerabilities identified had been closed by the end of the study.
Exploit windows collapse
The report sets this backlog against a sharp drop in the time it takes attackers to exploit a flaw. In 2018, the average time from a vulnerability’s disclosure to its first exploit was 2.6 years. According to the Zero Day Clock, it is now less than a day.
Black Duck points to AI as a major driver. The report cites Anthropic lab tests in which Claude Mythos Preview produced 181 working Firefox exploits under the same conditions in which its predecessor produced two. It also cites an AI-driven attack documented by Sysdig in February 2026, which reached admin-level access in eight minutes.
Black Duck expects capabilities comparable to Mythos Preview to become widely available among AI models within the next six to 12 months. The company has built what it calls a Mythos Readiness Framework, which measures how quickly organisations can act across four areas: visibility, coverage, prioritisation and remediation speed.
Remediation still on a monthly clock
The research found that even the fastest-fixed vulnerabilities, from static code analysis and open source scanning, took a median of 28 and 39 days to close. Those found through manual testing took 232 days, and those found through dynamic testing took 316 days.
The report also found that AI-assisted development and growing software complexity are adding to the volume of findings. Critical and high-severity issues rose from 13.1% of new findings in June 2025 to 16.8% in May 2026. In April 2026 alone, organisations added 10.4 million unresolved findings to their backlogs.
Prioritisation also appears to be breaking down under the volume. Critical open source vulnerabilities took a median of 58 days to close, more than twice as long as high-severity ones at 22 days. In Black Duck’s 2025 advisories, known exploits existed for 45% of high-severity, 39% of medium-severity and 31% of low-severity vulnerabilities. The report argues that severity scores alone are an incomplete guide to risk, especially as AI models can chain lower-severity weaknesses into serious attacks.
“Finding vulnerabilities is no longer the hard part,” said Greg Hughes, CEO of Black Duck. “Boards and regulators will judge organisations on time to fix, not on how many vulnerabilities they found.”
AI proposed for defence too
The report argues that human-paced remediation cannot close a gap that grows by millions of findings a month, and that more headcount will not solve it. It proposes three layers working together: deterministic testing for consistent, auditable results; AI-assisted triage to decide what to fix first based on exploitability, reachability and business context; and agentic remediation, in which AI generates and validates fixes that developers then review and approve.
The data suggests that automation is already making a difference. Organisations that scanned open source dependencies 10 to 29 times a month fixed vulnerabilities in an average of 30 days, against 66 days for those scanning less than once a month.
The findings are based on aggregated, anonymised telemetry from Black Duck customers. The company notes that, because these organisations invest in commercial security tooling, they may be more mature than the wider market.



