: The AI Vulnerability Race Is Not Coming. It Is Already Here.<\/span><\/a>\n\n\nGoogle disrupted the campaign before it launched. John Hultquist, chief analyst at the Threat Intelligence Group, put the significance plainly: “There’s a misconception that the AI vulnerability race is imminent. The reality is that it’s already begun. For every zero-day we can trace back to AI, there are probably many more out there.”<\/p>\n\n\n\n
That was Monday.<\/p>\n\n\n\n
By Tuesday, Microsoft publicly revealed MDASH, its multi-model agentic scanning harness, built by the Autonomous Code Security team alongside the Windows Attack Research and Protection group. MDASH orchestrates more than 100 specialised AI agents across an ensemble of frontier and distilled models to discover, validate, and prove exploitable vulnerabilities in complex codebases. It does not hand engineers a list of potential issues. It constructs triggering inputs, validates exploitability, and delivers confirmed findings. Sixteen of the vulnerabilities patched in this month’s Patch Tuesday were found by MDASH, not by human researchers. Four were rated critical. Two of those, a pre-authentication double-free in the Windows IKEv2 service affecting VPN infrastructure, and a use-after-free in the Windows IPv4 stack reachable by unauthenticated remote attackers, had CVSS scores of 9.8. A third critical flaw in Windows Netlogon, also unauthenticated and remotely exploitable, targets domain controllers directly. The researchers who reviewed the findings noted that the Netlogon vulnerability is wormable. A compromised domain controller is a compromised domain.<\/p>\n\n\n\n
Also on Tuesday, OpenAI launched Daybreak, a direct competitor to Anthropic’s Project Glasswing. Built on GPT-5.5 and the Codex Security agentic harness, Daybreak combines threat modelling, vulnerability discovery, patch validation, and dependency risk analysis into a single platform integrated into enterprise development pipelines. It is not yet fully public, but Cisco, Palo Alto Networks, Cloudflare, CrowdStrike, and Fortinet are already on board as partners. The race between Anthropic and OpenAI for this market is now explicit and public.<\/p>\n\n\n\n
Earlier in the week, the Risky Business security podcast ran a Between Two Nerds segment on why criminal organisations are structurally better positioned to adopt AI than legitimate businesses. The argument is sharp and worth sitting with: criminal enterprises have no compliance overhead, no governance approval cycles, no procurement processes, no legal review, and no board to brief. When a new AI capability becomes available, they can adopt it immediately and test it against live targets. The same capability that takes an enterprise security team six months to evaluate, procure, deploy, and operationalise can be in production use by a criminal group within days.<\/p>\n\n\n\n
Picus Security this week published a timeline that makes the asymmetry concrete. A fully automated attack chain starting from a known CVE reaches complete compromise in seventy-three seconds. The SIEM alert fires at one minute, after the attacker has finished. A Tier 1 analyst picks it up at minute five. A SOAR playbook gets triggered by hand at minute fifteen. A Jira ticket is filed an hour in. The patch goes out the following day, twenty-four hours after a breach that took seventy-three seconds to complete. The bottleneck is not the tooling. It is every human handoff in the response chain.<\/p>\n\n\n\n
That structural advantage is not theoretical. Google’s report this week also documented North Korean military group APT45 using AI to test and validate thousands of exploits at scale, Chinese state-linked operators experimenting with AI for vulnerability hunting and automated target probing, and a malware strain called PromptSpy that uses Gemini to autonomously navigate infected Android devices by interpreting on-screen activity and generating commands in real time.<\/p>\n\n\n\n
This is the picture: criminal and nation-state actors are already using AI offensively across the full attack lifecycle, from reconnaissance through exploit development to post-compromise operations. The major AI labs are racing to build defensive equivalents. Microsoft has deployed its own system internally and found sixteen vulnerabilities in Windows this month that human researchers had not. The defensive tooling is arriving, but it is arriving into organisations that are not yet ready to use it effectively.<\/p>\n\n\n\n
The more uncomfortable observation, which the Risky.biz discussion framed cleanly, is about the asymmetry of adoption friction. The organisations that need defensive AI tools the most, those with complex legacy estates, large attack surfaces, and constrained security teams, are also the ones with the most process overhead between a new tool appearing and that tool actually being used. The organisations with the least friction are the ones with the most to gain from the offensive application of the same capabilities.<\/p>\n\n\n\n
Microsoft’s Taesoo Kim, VP of Agentic Security, put the defensive argument precisely in the MDASH announcement: “AI vulnerability discovery has crossed from research curiosity into production-grade defence at enterprise scale, and the durable advantage lies in the agentic system around the model rather than any single model itself.” That last part matters. The model is not the moat. Any well-resourced actor, defensive or offensive, can access frontier models. The advantage sits in the pipeline, the orchestration, the institutional knowledge of where to look and what to do with what you find.<\/p>\n\n\n\n
That is an argument for investing in capability now, before the gap between offensive and defensive AI maturity widens further. It is also an argument for honest assessment of where your organisation currently sits. Not whether you have heard of these tools, but whether you have a plan for integrating AI-assisted vulnerability discovery into your security programme before your next Patch Tuesday.<\/p>\n\n\n\n
The race is not coming. The evidence from this week alone confirms it has started.<\/p>\n","protected":false},"excerpt":{"rendered":"
On Monday, Google confirmed the first criminal use of AI to develop a zero-day exploit.
\nOn Tuesday, Microsoft revealed an AI system that found 16 Windows vulnerabilities this month that human researchers had not found, including a wormable domain controller flaw.
\nAlso on Tuesday, OpenAI launched Daybreak, a direct competitor to Anthropic’s Mythos, making AI-powered vulnerability discovery a publicly contested market.
\nRisky.biz put something plainly this week that is worth hearing: criminal organisations are structurally better positioned to adopt AI than legitimate businesses. No compliance overhead. No procurement cycles. No board approval. When a new capability appears, they can test it against live targets immediately.
\nThe defensive tooling is arriving. The question is whether it can arrive fast enough, and into organisations that are ready to use it.
\nI have written about what this week actually tells us, why the asymmetry of adoption friction matters more than the technology itself, and what the honest question is for security leaders right now.
\nMore of my insights are available at auravere.com\/insights<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"closed","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"advanced_seo_description":"Google confirmed the first criminal AI zero-day this week. Microsoft's AI found 16 Windows flaws humans missed. The race is not coming. It is already here.","jetpack_seo_html_title":"The AI Vulnerability Race Is Already Here | Auravere","jetpack_seo_noindex":false,"jetpack_seo_schema_type":"","_kad_post_transparent":"","_kad_post_title":"","_kad_post_layout":"","_kad_post_sidebar_id":"","_kad_post_content_style":"","_kad_post_vertical_padding":"","_kad_post_feature":"","_kad_post_feature_position":"","_kad_post_header":false,"_kad_post_footer":false,"_kad_post_classname":"","_jetpack_newsletter_access":"","_jetpack_dont_email_post_to_subs":false,"_jetpack_newsletter_tier_id":0,"_jetpack_memberships_contains_paywalled_content":false,"_jetpack_feature_clip_id":0,"_jetpack_memberships_contains_paid_content":false,"footnotes":"","jetpack_post_was_ever_published":false},"categories":[22,29,24,23],"tags":[44,14,10,9,43,15,45,42],"class_list":["post-342","post","type-post","status-publish","format-standard","hentry","category-cyber-awareness","category-cyber-risk","category-information-security","category-security-leadership","tag-allinsecurity","tag-cissp","tag-cyber-security","tag-information-security","tag-patchtuesday","tag-security-leadership","tag-vulnerability-management","tag-zeroday"],"jetpack_featured_media_url":"","jetpack_sharing_enabled":true,"_links":{"self":[{"href":"https:\/\/www.auravere.com\/insights\/wp-json\/wp\/v2\/posts\/342","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.auravere.com\/insights\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.auravere.com\/insights\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.auravere.com\/insights\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.auravere.com\/insights\/wp-json\/wp\/v2\/comments?post=342"}],"version-history":[{"count":2,"href":"https:\/\/www.auravere.com\/insights\/wp-json\/wp\/v2\/posts\/342\/revisions"}],"predecessor-version":[{"id":367,"href":"https:\/\/www.auravere.com\/insights\/wp-json\/wp\/v2\/posts\/342\/revisions\/367"}],"wp:attachment":[{"href":"https:\/\/www.auravere.com\/insights\/wp-json\/wp\/v2\/media?parent=342"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.auravere.com\/insights\/wp-json\/wp\/v2\/categories?post=342"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.auravere.com\/insights\/wp-json\/wp\/v2\/tags?post=342"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}