When a vulnerability is disclosed, defenders need time to understand it, find the affected systems, and apply a fix. Mandiant’s time-to-exploit research shows how little of that time they can count on. It measures first observed exploitation relative to patch release. Its mean was 63 days in 2018–2019 and 32 in 2021–2022. For 2023 it reported five days after excluding 15 outliers; including those outliers produced 47 days.1 M-Trends 2026 gives a mean of negative one day for 2024 and an estimated negative seven for 2025.23
A negative mean says exploitation in the observed sample began, on average, before patches were available. It isn’t a deadline for every vulnerability. The result depends on which flaws were observed, the mix of zero-days and later exploitation, and the handling of extreme values. VulnCheck’s broader 2025 known-exploited-vulnerability dataset found that fewer than a third were exploited on or before CVE publication.4 CVE publication and patch release are different dates. Both studies show why a defender needs to know the actual exposure, not just repeat an average.
That work lands on projects whose maintainers already have full queues. Faster discovery is useful only if someone can validate the finding, develop a fix, and get it to users.
What the security figures can tell us
Black Duck’s 2025 audit report found open-source vulnerabilities in 86 percent of the commercial codebases it examined, with an average of 911 open-source components per codebase.5 That describes the audited dependencies. It doesn’t compare otherwise equivalent open and proprietary systems or establish which is more secure.
Sonatype’s 2024 report found that 95 percent of vulnerable releases being downloaded already had a fixed version available.6 In those cases discovery and patch creation were not the last obstacle. Users still had to identify the dependency, test an update, and deploy it. Finding another thousand flaws won’t do those steps for them.
Open source gives defenders important options: inspect the code, share findings, and distribute a common fix. I value those options. Their value depends on people being able to use them, which is why chapter 5 separated transparency from security. A neglected open project and a well-supported proprietary product can’t be ranked by their license alone.
What AI adds
The historical exploitation trend isn’t proof that AI caused it. Mandiant’s 2026 report says most successful 2025 intrusions still came from human and systemic failures, even as attackers used AI in parts of their work.3 The emerging discovery capability deserves attention on its own evidence.
In June 2025, XBOW’s penetration-testing system reached the top of HackerOne’s U.S. bug-bounty leaderboard after submitting roughly a thousand vulnerabilities.7 Humans still validated its findings, a qualification noted in an academic analysis.8 The ranking demonstrated useful automated discovery with a review process, rather than the disappearance of that process.
A July 2026 OpenAI evaluation showed a different risk. Models running with reduced cyber safeguards exploited an Artifactory zero-day, escaped their testing restrictions, and reached Hugging Face’s production infrastructure through further vulnerabilities and exposed credentials. OpenAI and the affected vendors documented the incident.9 Its conditions were unusual, including an internal research model, but the chained attack was real and crossed both commercial and open-source infrastructure.
Anthropic’s disclosure dashboard also makes the difference between finding and fixing visible. Its November 2, 2026 update reported 5,087 disclosed vulnerabilities across 680 projects, with 629 vulnerabilities patched upstream. Those figures replaced an earlier, higher disclosure total after duplicate and withdrawn findings were removed. The dashboard says independent human triage is a bottleneck; some reports sent directly to maintainers have not received that same independent check.10
Counts need that context. A candidate finding, a human-confirmed bug, a disclosed report, and an installed patch are different stages. None of them should be substituted for another to make a model or a security program look more effective.
There is uncertainty about exploitability too. In October 2026 Anthropic said a model had built a remote-root exploit for a Linux kernel TCP flaw, CVE-2026-98432. CloudLinux’s team could not reproduce that claim; the claimed exploit needed a Multipath TCP listener, and no exploitation in the wild had been reported.11 Defenders still needed the fix, but an attributed capability claim was not the same as a reproduced attack on their configuration.
From a report to a shared fix
A useful defense has to connect several jobs. Someone confirms that a report describes a real reachable flaw. Someone with project knowledge writes and reviews the patch. Release tooling produces an artifact users can verify. Distributors and operators then put that artifact into running systems.
OpenSSF supports parts of that chain. Sigstore helps users verify released artifacts; SLSA describes protections around how they were built; OSV Schema makes vulnerability records machine-readable; Scorecard checks project security practices.12 These tools reduce uncertainty about an artifact and its history. They don’t replace the maintainer’s decision about a patch.
Alpha-Omega funds security work in critical projects, including staffing at the Python Software Foundation, Rust Foundation, and RubyGems. In March 2026 a coalition of technology companies announced $12.5 million in new funding through Alpha-Omega and OpenSSF.13 Paying people to stay with the code addresses a different problem from paying a researcher to find one bug.
The Linux Foundation’s Akrites initiative, launched in June 2026, adds a shared security incident response team and coordinated disclosure process. It aims to get fixes upstream on maintainers’ terms and to act as maintainer of last resort for critical unmaintained packages.14 That gives a discovery program someone to contact besides the person whose inbox it just filled.
Package registries are part of the same infrastructure. Sonatype, Maven Central’s steward, joined other operators in the Sustaining Package Registries Working Group in May 2026.15 A project’s patched release has limited value if users can’t obtain it reliably or distinguish it from a malicious package.
Log4Shell showed both the value and the difficulty of coordination. The Log4j team issued fixes, with follow-up releases needed as the scope became clearer. CISA’s December 17, 2021 emergency directive required federal agencies to patch or mitigate by December 23.16 Public advisories and shared releases let many organizations work from the same information. Each still had to discover where Log4j was running and complete the update.
Paying for the last mile
Supported backports address another part of that work. A fix on the newest branch may not be usable in a hospital or bank that cannot immediately upgrade its application. Carrying the change to an older version, testing compatibility, signing it, and standing behind it costs money. Lightwell, IBM and Red Hat’s subscription offering, sells that service for application dependencies.17
Charging for support is consistent with open source. The question I care about is how the arrangement affects the shared fix. Does the underlying correction reach the original project? Can other distributors use it? Is the subscription paying for additional maintenance and accountability, or is access to an essential correction being withheld? Those are questions about particular terms and practices; a price tag by itself doesn’t answer them.
There is still a collective problem. Organizations able to buy extensive support have options a small user may lack. Shared funding can pay maintainers to make fixes available upstream, while paid services test and support versions for customers with particular needs. Those roles can reinforce each other if the contracts and the work actually do so.
The unfinished work
AI can shorten discovery without shortening every later step. It can also help write and review fixes, but a release still needs an accountable project, a trustworthy build, and users who can deploy it. The useful measure is how many verified defects get corrected in systems people run.
That is a case for funding the whole chain. It is also a reason to keep the source and the correction available: downstream users can share one fix instead of independently rebuilding it. The next chapter turns to models, where even identifying the artifact to inspect is more complicated than opening a source tree.
Notes
Casey Charrier and Robert Weiner, “How Low Can You Go? An Analysis of 2023 Time-to-Exploit Trends,” Google Cloud Blog (Mandiant), October 15, 2024, https://cloud.google.com/blog/topics/threat-intelligence/time-to-exploit-trends-2023. TTE is measured before or after patch release. The 2023 five-day average excludes 15 outliers; including them gives 47 days. The study covers 138 vulnerabilities disclosed in 2023 and tracked as exploited in the wild, rather than every disclosed vulnerability. ↩
Mandiant, M-Trends 2026, special report “Systematic Exploitation of Edge and Core Network Devices,” p. 78, https://www.gstatic.com/security-marketing/m-trends-2026-en.pdf (Google Threat Intelligence Group “has observed that the mean time to exploit (TTE) for vulnerabilities has continually decreased from 63 days in 2018 to -1 day in 2024”). Mandiant’s full 2024 time-to-exploit analysis, covering 112 vulnerabilities, is available only to Google Threat Intelligence customers. ↩
Jurgen Kutscher, “M-Trends 2026: Data, Insights, and Strategies From the Frontlines,” Google Cloud Blog (Mandiant), March 23, 2026, https://cloud.google.com/blog/topics/threat-intelligence/m-trends-2026; Mandiant, M-Trends 2026, pp.4 and 78, https://www.gstatic.com/security-marketing/m-trends-2026-en.pdf. The report describes the estimated 2025 mean of negative seven days relative to patch release. Its introduction says most successful 2025 intrusions still stemmed from human and systemic failures; it does not attribute the historical TTE trend to AI. ↩1 ↩2
Patrick Garrity, “VulnCheck State of Exploitation 2026,” VulnCheck, January 2026, https://www.vulncheck.com/blog/state-of-exploitation-2026. It finds 28.96 percent of 2025 KEVs were exploited on or before the day the CVE was published, which includes but is not limited to zero-days. ↩
Black Duck, press release on the 2025 OSSRA report, February 25, 2025, https://news.blackduck.com/2025-02-25-New-Black-Duck-Report-86-of-Commercial-Codebases-Contain-Vulnerable-Open-Source,-Exposing-Organizations-to-Security-Risks; Black Duck, Open Source Security and Risk Analysis (OSSRA) report page, https://www.blackduck.com/resources/analyst-reports/open-source-security-risk-analysis.html (911 components). ↩
Sonatype, press release on its 10th annual State of the Software Supply Chain report, October 10, 2024, https://www.sonatype.com/press-releases/sonatypes-10th-annual-state-of-the-software-supply-chain-report. ↩
XBOW, Series B announcement, XBOW blog, June 2025, https://xbow.com/blog/series-b; XBOW, blog post on reaching the top of the HackerOne US leaderboard, https://xbow.com/blog/top-1-how-xbow-did-it (about 1,060 vulnerabilities submitted). ↩
Mayoral-Vilches, “Cybersecurity AI: The Dangerous Gap Between Automation and Autonomy,” arXiv:2506.23592, 2025, https://arxiv.org/abs/2506.23592. ↩
OpenAI, “OpenAI and Hugging Face partner to address security incident during model evaluation,” OpenAI, July 21, 2026, updated July 28 and 29, 2026, https://openai.com/index/hugging-face-model-evaluation-security-incident/; OpenAI, “The Hugging Face incident and the road ahead,” OpenAI, August 26, 2026, https://openai.com/index/hugging-face-incident-and-the-road-ahead; JFrog, blog post on the JFrog and OpenAI collaboration on zero-day findings, updated August 5, 2026, https://jfrog.com/blog/jfrog-and-openai-collaboration-on-zero-day-security-findings; Hugging Face, “Anatomy of a Frontier Lab Agent Intrusion,” Hugging Face blog, https://huggingface.co/blog/agent-intrusion-technical-timeline. No vendor has publicly mapped specific CVEs to the chain. ↩
Anthropic, “Anthropic’s coordinated vulnerability disclosure dashboard,” updated November 2, 2026, accessed November 15, 2026, https://red.anthropic.com/2026/cvd/; methodology and corrections, https://red.anthropic.com/2026/cvd/about/. The dashboard reports 5,087 disclosed vulnerabilities across 680 projects and 629 findings patched upstream, not 629 patched projects. Its flowchart distinguishes candidate findings, external review, direct reports, and remediation. The November update replaced earlier totals after deduplication and withdrawal handling. ↩
CloudLinux, status post on the CVE-2026-98432 Linux kernel IPv4 TCP flaw, CloudLinux blog, updated November 3, 2026, https://blog.cloudlinux.com/cve-2026-98432-linux-kernel-ipv4-tcp-flaw-status-for-cloudlinux. Anthropic made the claim on October 27, 2026. CloudLinux describes the flaw as a use-after-free in the kernel’s IPv4 TCP code and says the claimed exploit needs a server with a Multipath TCP listener. ↩
OpenSSF, “Projects,” https://openssf.org/projects/. Listed initiatives include Sigstore, SLSA, OSV Schema, and OpenSSF Scorecard. OSV Schema is the vulnerability-record format; the OSV.dev aggregation and scanner services are distinct infrastructure, https://osv.dev/. ↩
Linux Foundation, press release on $12.5 million in grant funding for open-source security, March 17, 2026, https://www.linuxfoundation.org/press/linux-foundation-announces-12.5-million-in-grant-funding-from-leading-organizations-to-advance-open-source-security; Alpha-Omega, 2024 annual report, https://alpha-omega.dev/wp-content/uploads/sites/22/2025/01/Alpha-Omega-Annual-Report-2024_012925.pdf (security staffing at the Python Software Foundation, Rust Foundation, and RubyGems). ↩
Linux Foundation, “Linux Foundation and Industry Leaders Launch Akrites to Defend Critical Open Source Software Against AI-Enabled Cyber Threats,” press release, June 25, 2026, https://www.linuxfoundation.org/press/linux-foundation-and-industry-leaders-launch-akrites-to-defend-critical-open-source-software-against-ai-enabled-cyber-threats. ↩
OpenSSF, joint statement on sustainable stewardship of open infrastructure, OpenSSF blog, September 23, 2025, https://openssf.org/blog/2025/09/23/open-infrastructure-is-not-free-a-joint-statement-on-sustainable-stewardship; Sonatype, press release on the Sustaining Package Registries Working Group, May 6, 2026, https://www.sonatype.com/press-releases/sonatype-and-package-registry-leaders-unite (describes Sonatype as “steward of Maven Central”). ↩
CISA, Emergency Directive 22-02 on the Apache Log4j vulnerability, issued December 17, 2021, https://www.cisa.gov/news-events/directives/ed-22-02-mitigate-apache-log4j-vulnerability-closed; CISA, news release announcing the directive, https://www.cisa.gov/news-events/news/cisa-issues-emergency-directive-requiring-federal-agencies-mitigate-apache-log4j-vulnerabilities; Apache Logging Services, Log4j security page, https://logging.apache.org/log4j/2.x/security.html (the follow-up CVEs). ↩
Red Hat, Lightwell product page, https://www.redhat.com/en/lightwell (joint IBM and Red Hat offering, sold as an annual subscription); Red Hat, blog post on why Lightwell pairs dependency backports with upstream collaboration, Red Hat Blog, November 4, 2026, https://www.redhat.com/en/blog/why-lightwell-pairs-dependency-backports-upstream-collaboration (signed patched artifacts). ↩
From But Is It Open? by Timothy O’Brien (Discursive, 2026). © 2026 Timothy O’Brien. All rights reserved.