Insight 01
The $3 exploit
A verified exploit costs a median of $2.83, end to end. 90% cost under $5, and the priciest in the whole dataset was $14.63. The economic filter between “a vulnerability exists” and “a working exploit exists” is gone.
Vol. 1 · Q2 2026 · AI × Vulnerability
The Economics of
Vulnerability
Exploitation with AI
Evidence-based research on how artificial intelligence is changing the economics of exploitation, across open-source ecosystems, known-exploited catalogs, and the long tail of vulnerabilities no human has weaponized.
Published 17 JUL 2026n = 3,029Verified by LoginsoftTLP:CLEAR
72%
of analyzed CVEs are exploitable by the Exploit Harness.
The findings
Five numbers that redraw the AI-native cyber offense & defense balance.
01 · Cost
$2.83
Median AI cost-to-create-exploit-poc
AI-native cyber offense just got incredibly cheap.
02 · Speed
11 min
Median AI time-to-create-exploit-poc
The only thing that changes with AI is the speed of response.
03 · Coverage
72%
Quantro Exploit Harness created exploits for 72% of CVEs
Nearly three-quarters of CVEs analyzed could be exploited by Quantro’s Exploit Harness. The numbers will only climb from here.
04 · Compute
308K
Tokens to create and validate a working exploit
A complete end-to-end exploit fits inside the context window, with room for two more to spare in a 1M-token model.
05 · Remediation
14 min
Median AI time-to-create-remediation
It’s not all bad. What AI breaks, AI can also fix, just as fast.
The story, in three parts
Vulnerability exploitation just got cheap and fast.
One story, in three parts: what collapsed, why every warning system missed it, and the one signal that keeps pace.
I
Insights 1–4
Cost was the defense
What AI-native cyber offense now costs to weaponize a known CVE.
Part I →II
Insights 5–8
The instruments and metrics are obsolete
Why every signal we use to prioritize defense misses it.
Part II →III
Insights 9–10
A new signal, and the way forward
A measure that isn’t blind, and a defense that keeps pace.
Part III →Part I · Insights 1–4
Cost was the defense
For decades, the implicit defense of most vulnerabilities was cost. Writing a reliable exploit took a skilled human days or weeks. We measured what it costs an autonomous agent instead. The unanswerable became a dataset.
Insight 02
Not zero-day. Zero-minute.
Median time from a known CVE to a verified exploit is 11 minutes, with 90% landing inside half an hour. The gap between disclosure and weaponization has effectively closed.
Insight 03
Three-quarters, and climbing
The harness drove 72% of the dataset, 2,183 CVEs out of 3,029, to a verified working exploit. Treat that as a floor rather than a ceiling.
Insight 04
Not hands-off yet, but closing
Of 998 human-reviewed exploits, 77% needed no correction at all and verified exactly as generated. That number falls with every version of the harness.
Part II · Insights 5–8
The instruments and metrics are obsolete
If offense has collapsed, the signals we use to prioritize defense should have caught it. One by one, the catalog, the probability, the severity, the “no public PoC” comfort, they turn out blind to AI-driven exploitation.
Insight 05
What the CISA KEV catalog can’t see
89% of AI-exploitable CVEs, 1,933 of them, aren’t in CISA KEV. Past exploitation predicts nothing about what an agent can weaponize next.
Insight 06Centerpiece
Why predict, when you can prove?
73% of AI-exploitable CVEs carry an EPSS below 0.25, squarely in the deprioritize band. The vulnerabilities the harness couldn’t exploit score statistically the same, so EPSS can’t tell them apart.
Insight 07
Severity is not exploitability
Nearly a quarter of exploited CVEs, 500 in all, score below “High” on CVSS. Rated low, exploited anyway.
Insight 08
No proof-of-concept, no problem
For 450 CVEs no public PoC existed. The harness wrote one from scratch for a median of $3.00 and 11 minutes, barely more than it spent adapting an existing one.
Stop reading about it
Your EPSS, KEV and severity scores can’t see this.
So don’t rank by them. See what an AI attacker sees on your domain, or score a CVE you care about — both free, in minutes.
ai-recon · free scan
See what's exploitable on your domain
ai-xi · exploitability index
Is this CVE exploitable by AI?
Part III · Insights 9–10
A new signal, and the way forward
Every instrument on the dashboard is blind. So we built one that isn’t, and found that the same machine speed that broke offense open can be turned to defense.
Insight 09
The only score that matters
The AI Exploitability Index (AI-XI) scores each vulnerability 1–5 by how hard the agent had to work. Cost and time rise cleanly with it, and even the hardest solved tier costs under $4.
Insight 10
What AI breaks, AI can fix
The harness generates a remediation in a median of 14 minutes, about as fast as it builds the exploit. To fight AI-native cyber offense, you need AI-native cyber defense.
Projects
2 active
Turn the research on your own stack.
Two free tools from the VRL research team. Scan your domain for AI-exploitable exposure, or score any CVE by whether AI can exploit it.
ai-xi
Exploitability index
Is this CVE exploitable by AI, and how hard? Paste a CVE and an autonomous agent builds and verifies a working exploit, then returns an index from 1 (easy) to 5 (hard) with time and cost to exploit — median 11 minutes, $2.83. Never the exploit code.
Paste a CVE to score it
ai-recon
Attack surface
See what's exploitable on your domain — free. Point the agent at a domain, org, or repository and it enumerates exposed services, packages, and versions, then flags where our dataset of verified exploits already has a working PoC for what you're running.
Enter a domain to scan free