The disclosure pipeline is triaging itself: 20,700 new CVEs, 10% more exploitation

Four load-bearing parts of the vulnerability disclosure pipeline changed how they operate in the second quarter, and each change removed coverage rather than adding it ^[3][4][5][6]. According to Beazley Security's Q2 2026 quarterly threat report, that retrenchment happened while disclosed vulnerabilities grew 36% to more than 20,700 and confirmed exploitation in the wild grew by only 10% ^[1][2].

The two figures are the argument. Volume is arriving faster than the institutions that process it can absorb, and almost none of the extra volume is showing up as attacker capability. Beazley attributes the surge to agentic AI producing reported vulnerabilities at scale ^[7], and reports that Anthropic's Mythos model was briefly restricted from foreign access by a US export order in June before the restriction was lifted the same month ^[8].

The responses are worth reading as a set. NIST has stopped enriching every CVE it receives ^[3]. HackerOne paused new submissions to its Internet Bug Bounty program ^[4]. Pwn2Own rejected contestant applications for the first time in its history ^[5]. Cisco rebuilt its disclosure model and now bundles multiple flaws under a single CVE, which Beazley describes as a departure from how CVE IDs are meant to work ^[6]. Three of those are queue management. The fourth is different in kind: if one identifier can cover several defects, then counting CVEs is no longer counting bugs, and any remediation SLA written in terms of CVE IDs quietly changes meaning without anyone editing the policy. The enrichment that NIST is no longer doing does not disappear either. It moves to whoever consumes the feed, which in practice means the vulnerability management team that was already behind.

Beazley's own numbers show what triage at this volume looks like. Of roughly 5,600 high-risk CVEs disclosed in the quarter, Beazley Security Labs issued 21 advisories, up 40% from Q1 ^[9]. That is about one advisory per 267 high-risk CVEs ^[1], and the high-risk pool itself is only around 27% of everything disclosed ^[2]. The 36% growth rate implies roughly 15,200 disclosures in Q1 ^[3], so the quarter added something like five thousand new records to the pile while the vendor lab watching it raised its signal count from 15 to 21 ^[4]. The filtering ratio is not a marketing point. It is a measurement of how little of the feed is actionable.

Meanwhile the entry points did not move. Beazley reports that 67% of ransomware intrusions began with compromised credentials against exposed VPN and RDP, and another 14% with malware delivered by SEO-poisoned installers ^[10]. What changed was after entry: a growing share of affiliates skipped encryption entirely for data theft and extortion, a pattern Beazley saw from Inc Ransomware, Brain Cipher and Pear affiliates ^[11]. For BEC, attackers increasingly abused Microsoft's device code sign-in flow, in which the victim enters an attacker-generated code on a genuine Microsoft page, so the login and the MFA prompt are real and the attacker collects the session token ^[12].

Watch whether other large vendors copy Cisco's bundling, because that is the change that breaks downstream counting rather than merely delaying it ^[6]. Watch whether HackerOne reopens IBB submissions with new intake rules ^[4]. And watch the Q3 ratio: if disclosures grow another third and exploitation again grows a tenth, the case for spending on identity at the VPN and RDP edge gets stronger, not weaker ^[1][2][10].