Nov 10, 2026
6 min read
2 views

A pre-registered count of prompt-injection records in the NVD finds a clear rise in the rate, an inconclusive test for a step at the agentic-product wave, and a record that two reporting organisations mostly write.
Glasp Research. Keisuke Watanabe, Glasp Inc.
We counted the vulnerability records that describe prompt injection, the attack where text an AI model reads is treated as an instruction. We found 162 in the National Vulnerability Database (NVD) for CVEs published between 1 January 2023 and 31 October 2026. Their share of all CVEs has risen. Whether the rise has a step at the time agentic products arrived, our test could not say.
We pulled every CVE published in that window from the NVD API on 8 November 2026. That was 213,944 records. Of these, 7,179 carry the status Rejected and were excluded, leaving 206,765.
The set of interest was fixed before we pulled anything: records whose English description matches the case-insensitive pattern prompt[ -]?injection, plus any record tagged with the weakness CWE-1427, which is the weakness type for improper neutralization of input used in prompts to language models. The pattern found 162 records. The CWE tag found none. We checked that zero before accepting it: the string "CWE-1427" appears once in all the raw data, inside the free text of a description, and the weakness field itself, which we confirmed was being read correctly, never lists it. NVD simply did not use that tag in this window.
We drew 100 of the 162 records at random and had two AI models label each one, blind, against a written rubric: is there a language model in the loop, and does attacker-controlled text in what it processes cause unintended behavior? There was no human label, and this is the main weakness of the check. The two models agreed on 99 of 100 records (Cohen's kappa 0.85). The single disagreement was settled by a third read and counted against the record.
The result was 96 yes, 1 no and 3 unclear, so label precision is 96%, with a 95% interval of 90.2 to 98.4. Scaled to 162 records, that is about 155.5 confirmed, with a range of roughly 146 to 159. The low end of that range sits under the 150 we had set as the minimum for going on, and we say so rather than rounding it away.
The raw count grew from 4 records in the first quarter of 2025, to 14 in the fourth quarter of 2025, to 39 in the third quarter of 2026. But the number of CVEs grew too, from 6,975 in the first quarter of 2023 to 36,954 in the third quarter of 2026, so the plan was to compare rates per 10,000 CVEs.
That rate was 0.00 to 1.40 across 2023, then 0.00 to 8.14 across 2024. In 2025 it ran from 3.35 to 10.95. It reached 15.84 in the first quarter of 2026, and was 10.63 and 10.55 in the second and third quarters. So the rate rose, then levelled off at around 10 or 11 in recent quarters. Seven of the 15 quarters have fewer than 5 records, so early quarters are noisy.
The plan named one question in advance: was there a step in the rate at the wave of agentic products in 2025? We fitted an interrupted time series, a count model with total CVEs as the exposure, at 1 July 2025, and at 1 January 2025 and 1 October 2025 as sensitivity checks. We committed to reporting all three, not the best.
The step estimate points in different directions depending on the date. The level-shift ratio was 0.42 for the January break (interval 0.17 to 1.03), 1.22 for the July break (0.59 to 2.56) and 1.17 for the October break (0.61 to 2.27). Every interval includes 1, which means no step is shown. The slope change was below 1 at all three dates (0.66, 0.78 and 0.73), meaning the rate grew more slowly after each break than before it. The January fit rests on only 20 events before the break. We pick no break as the right one: the test is inconclusive and depends on the break date.
The records are concentrated. The GitHub CNA issued 65 of the 162, which is 40%. VulnCheck issued 40, or 25%. The MITRE CNA issued 20. The top three together hold 77.2%.
All 40 VulnCheck records were published in 2026, and they are 37% of that year's 108 records, so part of the 2026 level reflects one organisation's practice. In 2025, GitHub alone was 23 of 34. Our stop rule for one organisation above 50% overall was not triggered, but a change in one organisation's habits could still move the series.
Of the 162, 80 (49.4%) have no vendor in the first product entry, and the missing share is mostly 2026, so vendor rankings would mislead and we do not give them. Of the 151 records with a CVSS v3 score, 71.5% were High or Critical. We did not pull the severity mix for all CVEs, so we cannot say whether that is high or typical.
This is not evidence that there are more flaws. Over this period more products may have added language-model features, and more bug-bounty programs and CNAs may have begun to report them, so a rising count can be a rising habit of reporting. We make no causal claim, and nothing here is an effect of agentic products.
Recall was not measured. The set is a regex on the description, so any record about prompt injection that uses other words is missing. Precision tells us how many of the 162 are right, not how many are absent.
The labels came from AI models and no human checked them; a spot check of 15 rows was also done by an AI session, not a person. Quarterly counts are small, and the model has no seasonal terms and does not model the jump in total CVEs in the third quarter of 2026. October 2026 is a partial quarter, with 23 records, outside the main analysis.
A conflict of interest: the analysis was done with Claude, made by Anthropic, and at least three of the 162 records concern Anthropic products (three have "anthropic" as the first listed vendor, and half the records list no vendor).
We wrote the plan, the set definition, the break dates and the stop rules before pulling any data. There were three stop rules: precision under 70%, an estimated confirmed count under 150, and one CNA above 50% of the set. None was triggered, though the second was close at the low end of its range. We changed nothing in the plan afterward, apart from repeating requests when NVD connections ended early.
Data came from the NVD API, with no key and a pause of at least 7.5 seconds between requests.
https://nvd.nist.gov/developers/vulnerabilities
The weakness definition is on the CWE page.
https://cwe.mitre.org/data/definitions/1427.html
Two earlier studies are related. arXiv 2604.04288 analyzed 295 GitHub advisories selected by keyword and gives no time trend. arXiv 2605.28893 studied 205 curated vulnerabilities and has no year-by-year series. Our contribution is a quarterly series from one database with a measured label precision, not a replacement for either.
This article was researched and written by Claude (an AI model by Anthropic) in a Glasp research run. No Glasp user data was used. Checks, 8 November 2026: a separate AI review compared every figure in this text with the result files, and the corrections it asked for were made. No human has checked these figures.