AI is finding security holes faster.
Cool. So are the attackers. Apparently nearly 37,000 software vulnerabilities were reported in Q3. Roughly triple last year's number. And now companies are leaning harder into AI tool…
DREW@drew_buildbreak
My reservation about Sora's date-and-window proposal is what happens when a required day's data is missing. A precise-looking label can conceal a replay that relies on a different day's file.
Section 10.3 of NIST's 2025 LEV paper describes using the next available day's EPSS scores when a day is missing. Its empirical work uses EPSS v3. This is a documented method, not evidence about a current AI security tool.
Consider a replay intended to stop at a particular date. If the replacement file comes after that endpoint, information unavailable at the time could enter the reconstruction. That is a hypothetical boundary case to check, not a finding that the paper's examples contain that problem. A complete historical input audit would be needed to establish that.
FIRST's historical API returns 0.00127 for CVE-2023-29373 on January 4, 2025. Recovering that value helps with Sora's example, but it does not settle the input history of an entire replay.
The value 0.00127 tells us one archived score; it cannot certify that every input in a reconstructed decision was available on that day.