TIPS #44: The Vulnpocalypse and the Remediation Gap
Shane Shook
October 27, 2026
- Blog Post
- TIPS
Issue: AI is collapsing the cost of finding, exploiting, and reporting vulnerabilities, and remediation is falling behind.
Traditionally, vulnerability management and bug bounty programs were built on the assumption that finding real vulnerabilities is hard, which kept the number of findings manageable. Security teams could scan, score, and patch vulnerabilities on a reasonable timeline, and bounty programs could pay for validated reports.
AI-assisted vulnerability discovery is changing the equation and creating an industry-wide remediation gap.
AI-assisted discovery is driving record CVE volume
After CVEs more than doubled between 2020 and 2024 and hit a historic 48,244 in 2025, 2026 is already another record-setting year for vulnerability disclosures.
In February, the Forum of Incident Response and Security Teams (FIRST) issued a median forecast projecting 59,400 CVEs in 2026. By June, FIRST raised its projection to 66,000 CVEs, citing AI-assisted discovery as a key driver.
Preliminary data from Q3 2026 indicates that vulnerability disclosures have only accelerated since then. A search of NIST’s National Vulnerability Database (NVD) as of October 20, 2026, shows approximately 37,700 CVEs published in Q3– more than in the entire first half of the year- bringing the 2026 total to over 72,000 so far.
AI-assisted discovery is driving much of this volume. Researchers, defenders, and attackers are using AI to discover vulnerabilities more cheaply and quickly. One of the most publicized examples came in April when Anthropic announced Project Glasswing, which gave around 50 organizations including Apple, Cisco, Microsoft, Google, and the Linux Foundation access to Mythos Preview, an unreleased frontier AI model that could find and exploit software vulnerabilities on its own. In the first month of the project, Glasswing partners found more than 10,000 high- or critical-severity vulnerabilities in widely used software and several reported finding bugs more than 10 times faster than before.
Six months later, the impacts are felt well beyond Project Glasswing. Forgepoint portfolio company Hackuity recently reported that across 20 of its largest enterprise customers, the volume of vulnerabilities to manage grew 4.4-fold between January and October 2026, and the monthly inflow of new vulnerabilities rose 5.7-fold after Mythos-class AI models arrived in April.**
Remediation lags
Security teams now face a flood of findings from scanners, tools, and industry research. Yet their capacity to prioritize, verify, and patch critical and exploitable vulnerabilities has not kept pace, leading to a growing remediation gap.
The proof is in the data. Through April 2026, actionable exploitability remained relatively flat as CVEs surged,* yet vulnerability exploitation is the top access vector for attackers according to both the 2026 Verizon Data Breach Investigations Report (DBIR) and the Mandiant M-Trends 2026 report. In other words, it’s becoming harder for security teams to find the small number of exploitable, high-risk vulnerabilities relevant to their organizations amidst the high volume of CVEs.
Attackers often exploit vulnerabilities faster than vendors can issue patches, while defenders take longer to remediate known exploitable vulnerabilities. Mandiant’s M-Trends 2026 report found that the mean time to exploit a vulnerability was an estimated negative 7 days in 2025, meaning exploitation began a full week before a patch existed on average. At the same time, the 2026 Verizon DBIR found that the median time to fully patch rose from 32 to 43 days and that organizations fully remediated only 26% of Known Exploited Vulnerabilities (KEVs) in 2025, down from 38% the year before.
Vulnerability programs face challenges
The capacity gap extends to the shared systems that score vulnerabilities and transform outside research into patches. Two of the clearest examples of this can be seen in recent changes to NIST’s NVD and several large bug bounty programs.
On April 15, 2026, NIST fundamentally changed how its NVD- the U.S. government’s repository of standards-based vulnerability data- operates, citing a 263% increase in CVE submissions between 2020 and 2025. NIST now only prioritizes enrichment- adding key details like severity scores and affected product lists- for CVEs in CISA’s KEV catalog, software used by the federal government, and critical software as defined by Executive Order 14028. All other CVEs are labeled “lowest priority” and are not scheduled for immediate enrichment. As a result, many new CVEs now reach scanners without the severity scores and product mappings those tools depend on, and some will undoubtedly be exploited before they reach the KEV catalog.
Bug bounty programs are facing similar challenges. AI tooling has made it much easier to find bugs and generate bounty reports, yet each report still needs a skilled expert to reproduce, validate, and close the vulnerability or exploit. The cost of bug bounties has shifted from discovery to validation, and many programs simply can’t keep up with the volume of reports. Several notable bug bounty programs have restructured, paused operations, or ended this year:
- curl ended its paid HackerOne bounty in January after a deluge of low-quality, AI-generated reports.
- HackerOne paused Internet Bug Bounty program submissions in March, stating that AI-assisted research had shifted the balance between findings and remediation capacity in open source. Node.js, which relied on the program’s funding, stopped paying rewards as a result.
- GitHub restructured its bug bounty program in July, creating a lower-paying public tier and a higher-paying invitation-only tier, citing a growing backlog of low-effort, AI-generated reports. The company also added a track-record requirement for new researchers.
The changes to the NVD and bug bounty programs mean that security teams receive less context on vulnerabilities, while software vendors face more barriers to validate and fix reported flaws. *Preliminary data from Q3 2026 shows the same trend, with one emerging difference: CISA added 100 vulnerabilities to its KEV catalog in Q3, up from 71 in Q1 and 75 in Q2. So, while KEV additions are still a small and shrinking share relative to new CVEs (about 0.27% in Q3 2026, down from 0.42% in Q3 2025, by our calculations based on the KEV catalog and the NVD database), the raw number is now rising, too.
**Anonymized, aggregated Hackuity customer data from October 2026, shared with Forgepoint.
Impact: Critical vulnerabilities stay open longer and attackers exploit the gaps.
When reporting, patching, and prioritization lag, critical vulnerabilities go unpatched for longer and attackers take advantage of exploitable gaps. The following case studies show how this dynamic is playing out.
Cisco firewall vulnerabilities and patching delays (September 2025-April 2026)
On September 25, 2025, Cisco disclosed that a threat actor linked to the ArcaneDoor state-sponsored espionage campaign was exploiting zero-day vulnerabilities (CVE-2025-20333 and CVE-2025-20362) in its Adaptive Security Appliance (ASA) and Firewall Threat Defense (FTD) products. The company released patches for the impacted products on the same day, and CISA quickly issued Emergency Directive 25-03 ordering federal agencies to identify and patch affected devices within days. Yet remediation lagged for weeks:
- On September 30, The Shadowserver Foundation found that nearly 50,000 internet-exposed Cisco firewalls were still unpatched.
- On November 12, CISA found that some federal agencies had reported devices as “patched” when they were still running vulnerable versions.
Patching once also wasn’t enough: in April 2026, CISA updated its directive after ArcaneDoor developed a new persistence mechanism that survived the September fixes and expanded its targeting to any device running ASA or FTD software.
In this case, a sophisticated threat actor targeted two zero-day vulnerabilities, and despite clear vendor disclosures backed by CISA mandates, remediation lagged for weeks across public and private organizations. It’s a good example of how exploited flaws in perimeter devices- some of the most targeted assets in the enterprise- compete for attention with tens of thousands of lower-risk CVEs and can stay open for weeks or months. This campaign also shows how a device reported as patched is not necessarily remediated, and why fixes need ongoing verification as attackers adapt.
Apple bug bounty restrictions (June 2026)
In June, Apple introduced a bug report cap and a 30-day cool-off period after receiving a surge of AI-generated submissions. Following the bug bounty program changes, cybersecurity startup Bynario used ChatGPT to find more than 50 potential macOS bugs in three weeks, including a privilege escalation exploit chain that could give an attacker full control of a Mac. However, the company was unable to submit the exploit because it had reached Apple’s new report cap. Apple began reviewing Bynario’s findings only after the Financial Times reported the story in August.
In this case, a control meant to reduce noise in a bug bounty program blocked a critical report, leaving an exploitable flaw unaddressed until media coverage prompted a review. Apple’s goal of protecting its triage team from low-quality reports was reasonable, but the mechanism was imprecise: a volume-based report limit does not distinguish noise from signal.
Microsoft’s record Patch Tuesday (September 2026)
On September 8, 2026, Microsoft patched a record 974 vulnerabilities across Windows, Office, SQL Server, and other products, bringing the firm’s 2026 total to over 2,600 patched vulnerabilities. More than 110 of the vulnerabilities were rated critical, and several unexploited flaws in the release scored as high as 9.8 on the Common Vulnerability Scoring System (CVSS).
The release included two Windows zero-days under active exploitation: CVE-2026-85880 in Windows Advanced Local Procedure Call (ALPC) and CVE-2026-81963 in the Windows Update Stack. Both exploits are based on local privilege escalation bugs that enable an attacker with a foothold to gain SYSTEM-level control of the machine, and both carry a CVSS score of 7.8. CISA added the two exploited CVEs to the KEV catalog the same day.
In this case, two exploited vulnerability patches were released alongside nearly 1,000 other patches in a single month by a single vendor. It’s a clear example of the flood of CVEs and patches from the defender’s perspective: few security teams can manually research, test, and deploy that many patches at once. Most fall back on triaging by severity, which in this case would lead them to patch the unexploited 9.8 CVSS vulnerabilities first, leaving time for attackers to exploit the 7.8 CVSS vulnerabilities.
Severity scores are an important piece of the puzzle, but they must be evaluated in context: whether a flaw is being exploited, where an affected system is in the business, and whether attackers are already inside the network. As NIST notes, “The Common Vulnerability Scoring System (CVSS) is a method used to supply a qualitative measure of severity. CVSS is not a measure of risk.”
Action: Define the signals, reduce the noise, and close the remediation gap.
To combat the flood of vulnerabilities, companies need to identify which vulnerabilities matter to their business, fix the most important issues first, and reduce their overall risk profile.
1) Model dependencies as a graph
Start by defining the vulnerability and risk environment in the organization. Think of enterprise information systems as a graph. Endpoints such as users, devices, applications, and APIs are nodes, and the interactions between them are edges. Each edge has a weightthat reflects its importance to the business, and the weight changes depending on business requirements and circumstances.* For example, a payments API matters more during quarter-end close, while a remote access gateway matters more when employees work remotely.
Kept current, the graph gives a real-time picture of the key processes and resources that need protection and exposes single points of failure. A CVE on a node with no priority edges is noise, regardless of its severity score, while a CVE on a node connected to a heavily weighted edge is signal, especially if there is no redundant path.
*Note: This model is adapted from Mission Dependency Modeling, an approach MITRE researchers developed and articulated for military cyber situational awareness in 2016.
2) Shrink the attack surface
Next, switch images with ones that have the maximum number of security bugs patched and remove unnecessary components to reduce noise without losing key signals. Most containerized applications ship with packages, libraries, and binaries that are never executed, each of which introduces new CVEs. Security teams can eliminate CVEs by deploying clean, curated images then removing unused components, which when combined shrink the attack surface and the number of security threats prior deployment.
RapidFort replaces images with minimal, curated images, then optionally profiles what runs in production and removes unused components, delivering near-zero CVE container images without code changes.
“Security teams can find vulnerabilities, but they often can’t keep up with fixing them. Start by replacing vulnerable container images with curated near-zero CVE images, then use runtime profiling to identify and remove unused components. This approach eliminates vulnerabilities and continuously reduces exposure without code changes, helping teams close the gap between finding and fixing.”
Mehran Farimani CEO and Co-founder, RAPIDFORT
3) Prioritize remediation for high-risk assets
Remediation should start where exploitability and business impact overlap: exploitable flaws on priority edges and nodes. Combine exploitability signals such as CISA’s KEV, threat intelligence, and FIRST’s Exploit Prediction Scoring System (EPSS)- which estimates the likelihood a published CVE will be exploited in the next 30 days- with each asset’s position and weight on the organization’s dependency graph. Internet-facing edge devices on priority paths should be at the top of the list. It’s also critical to coordinate remediation across security, IT, and engineering teams, and to confirm that fixes were applied.
Hackuity’s Vulnerability Operations Center (VOC) platform helps security teams prioritize and remediate the most critical vulnerabilities and business risks. Its AI-powered technology aggregates findings across the security stack, adds threat, asset, and business context to each CVE, and automates remediation workflows through to verified closure so teams can focus on the risks that matter most.
“Detecting vulnerabilities isn’t the challenge. Managing their volume is. AI-driven discovery now outpaces any human’s capacity to respond, and severity scores alone can’t tell teams what to fix first. Enterprises need to automate prioritization based on real-world risk and drive remediation at machine speed.”
Patrick Ragaru CEO and Co-founder, Hackuity
4) Confirm what’s exploitable and if it’s being exploited
While prioritization shows security teams where to look, validation addresses whether an attacker can reach and exploit a vulnerability, and if any attackers already doing so. This must be done continuously: since exploitation often starts before a patch exists, security teams can’t wait for the next scan cycle.
On the exploitability side, Bishop Fox’s Continuous Security Testing (CST) services help security teams continuously discover the organization’s external attack surface, confirm which exposures are exploitable, and test the environment against emerging threats.
“As the cost of producing plausible output falls to zero, the cost of validating it rises, and the only thing that pays it down is human judgment. AI can surface more potential vulnerabilities than ever, but an expert still has to confirm which ones attackers can actually reach and exploit.”
Vinnie Liu CEO and Co-founder, Bishop Fox
On the active exploitation side, Qevlar AI’s Vulnerability Exploitation Hunter translates newly disclosed CVE data into threat hunt queries and searches environments for signs of exploitation from the day of disclosure. Its CVE Exploitation Intelligence Exchange lets SOC and vulnerability teams share results and work from the same data.
5) Treat bug bounty and disclosure reports as a signal, and integrate them with internal vulnerability management processes
Despite the flood of low-quality, AI-driven submissions and reports, bug bounty programs still matter because they help organizations identify flaws they can’t find on their own. But as bug bounty pioneer Katie Moussouris has pointed out, these programs only work when backed by mature internal vulnerability management processes that prioritize, validate, and patch bugs; organizations can’t just bounty their way to being secure, nor can they only patch their way there.
Bounty reports can be a valuable signal if programs are designed to reward real impact and findings flow into the vulnerability prioritization process.
If you run a bug bounty program:
- Route validated reports into your vulnerability management workflow alongside internal findings
- Pay for demonstrated impact, not report volume
- Require a working proof of concept
- Weight rewards by business impact on priority processes rather than CVSS alone
- Give researchers with a strong track record a faster reporting path
- Make sure submission caps can’t block a critical report
If you rely on vendors’ bug bounty programs:
- Their backlog is your exposure. Ask software suppliers how they handle AI-generated reports and how quickly they validate and patch critical findings
- Feed vendor advisories and patches into the same prioritization process as your internal findings
- While a vendor fix is pending, use exposure testing, exploitation hunting, and compensating controls to cover gaps