The Signal in the Noise: Why 98.6% of Vulnerabilities Don’t Matter and the 1.4% That Will End Your Business
By Bruce Johnson, Vice President, Solutions
I just came back from the Gartner® Security & Risk Summit 2026, and there were a lot of observations that struck me as being relevant to the challenges we’re trying to chase here at TekStream. One of the sessions I particularly liked was from Craig Lawson, titled, We Can’t Patch Our Way Out of Threat Exposure Debt; it was clarifying. The key takeaway for me, based upon Lawson’s research, was that every year, the vulnerability universe expands, not relevatory in and of itself. In 2025, security teams received notification of more than 45,000 new CVEs, 5,475 rated Critical, 13,051 rated High, and a medium-severity backlog that has grown nearly 150% since 2021 alone. OSINT feeds are broader, threat intelligence subscriptions are richer, and the attack surface our organizations are exposing to the internet grows every quarter with each new SaaS integration, cloud workload, and machine identity. Kind of explains why patch Tuesday has leaked into Wednesday.
What was interesting was that, of the 12,362 CVEs that had a known exploit available in 2025, only 606 were confirmed exploited in the wild.
That’s 4.9% of vulnerabilities with an available exploit. Less than 1.4% of all reported CVEs.
I won’t say that vulnerabilities are a distraction, but it seems equivalent to going to the doctor for a peanut allergy and instead of being prescribed an epi-pen, you’re told to eat your vegetables. We view these AI generated threats as more on the existential side, given the level of preparedness to respond with “human on the loop,” machine-speed coordinated defense.
So, if the exploitable surface is actually small and relatively predictable, the question isn’t how many vulnerabilities you can patch. The question is whether your security program is pointed at the right ones, before threat actors find them in your environment. We are shifting our focus from overall risk to exploitable risk, from compliance to focused threat protection.
Most programs aren’t.
In fact, based upon my highly scientific, anecdotal, casual conversations with CISOs, everyone I’ve spoken with says they aren’t seeing AI threats against their environment. Not to be a contrarian, but everywhere we’ve deployed our deception engineering solution, we’ve seen verified daily activity that incorporates AI-driven tells. These indicators are identified not just by velocity and sophistication, but across a number of qualifying dimensions that point to AI involvement, in some cases down to the specific LLM being used.
The Debt You’re Carrying Without Knowing It
Threat exposure debt is technical debt your organization carries that can be actively exploited by threat actors. Unlike software debt, you don’t get to defer it in a sprint backlog. It compounds while your team is watching something else.
The dominant response to this debt, and full disclosure, the approach that we’ve traditionally chased with TDIR—the SLA-based, severity-scored, compliance-driven patching and TDIR approach—is fundamentally misaligned with the threat. Even the UI at that level is inadequate to respond to a coordinated attack. Instead of approving an AI plan that outlines a composite threat, analysts are doom scrolling through IOCs like it’s an endless late night Instagram feed. Compliance frameworks currently treat every Critical CVE as roughly equivalent. Threat actors do not. They are starting to exploit the specific intersection of a known CVE, your specific platform configuration, your specific identity topology, and your specific detection gap. That intersection is unique to your environment, and no generic threat feed tells you where yours is.
The 2024 data makes the misalignment a little less abstract: 230 CVEs were exploited on or before the day of disclosure, zero-days at record levels. The gap between “CVE published” and “exploit in the wild” has collapsed to hours in the most dangerous cases. No SLA-driven patching program is designed for that timeline. You cannot compress your way to zero exposure. What you can do is shrink your relevant exposure, the intersection of what’s in your environment and what attackers are actually pursuing.
That requires something compliance frameworks cannot provide: customization at the level of your actual environment, your actual threat actors, and your actual business processes.
Attack Paths Are Predictable. Until They Aren’t.
Here’s the counter-intuitive good news: despite the volume of CVEs, the population of vulnerabilities that actually get exploited over any trailing ten-year window is remarkably constrained. Approximately 500 CVEs per year. Roughly 3,900 in the cumulative decade ending 2025. Advanced Persistent Threat actors follow documented playbooks. MITRE ATT&CK catalogs the techniques. The attack paths that most commonly threaten a mid-enterprise Splunk environment, a healthcare payer, or a federal contractor are known, enumerable, and, to a trained threat intelligence function, largely predictable.
That predictability is why proactive defense has historically been feasible, and why a prioritized, threat-driven approach to vulnerability management outperforms compliance checklists by orders of magnitude. That was another consistent theme at the Gartner conference, applying the “shift left” concept to becoming more proactive in our approach to threats. Particularly important because AI is beginning to erode predictability.
Checkpoint’s session at Gartner Summit 2026 demonstrated a five-step multi-agent attack: a support chatbot was compromised, which manipulated a product management agent, which injected into a cloud code agent, which passed a malicious commit through code review, which shipped to production. Five independently reasonable security controls saw only fragments of the chain. None caught it. No human analyst correlating alerts from five systems in real time could have caught it either.
This is the cross-domain correlation problem, familiar from the XDR era for human-actor attacks, now restated for AI agents. And it compounds fast. When an attacker establishes a foothold in your environment, the average breakout time is approximately ten minutes. Your mean time to detect, industry-wide, is still running around seventy days. That asymmetry doesn’t close through headcount. It closes through AI. It’s tricky, and not only do we have to detect super-fast but longer dwell times with unpredictable exploits are going to be a part of the challenge when velocity is classified as a principle indicator of AI attacks.
You Need an AI to Fight an AI
The security industry is in the earliest phase of an architectural transition that mirrors what happened with endpoint protection. The antivirus era was signature-based, stateless, reactive. EDR introduced contextual runtime detection, behavior over signatures. We are now at the threshold of what Gartner calls the “XDR for AI” moment: cross-domain, cross-agent correlation that no human-speed, alert-by-alert SOC can replicate.
The numbers support the urgency. Verizon’s DBIR reports a 31% increase in vulnerability exploits attributable in part to AI-enhanced attack chains. AI-enabled adversaries don’t just move faster, they attack more vectors simultaneously, degrading the assumption that human SOC triage can keep pace. Only 20% of cybersecurity leaders currently report AI as highly beneficial to their security operations. The gap between what AI-empowered attackers can do and what most AI-assisted defenders can actually do is real, and it’s widening.
The answer is not more analysts. The answer is not more alerts. The answer is an AI operating continuously at the intersection of your specific threat landscape, your specific environment, and your specific detection posture, with human operators to govern, validate, and refine every action.
That is exactly what TekStream built Cosmos to be, because we feel it’s not that using AI is important, it’s how you’re using AI that counts.
Cosmos: Differentiated Where It Matters
Cosmos is not a SIEM. It is not a threat feed subscription. It is not a compliance dashboard. It is an autonomous cyber defense framework purpose-built to shrink the distance between the threat landscape and your specific environment, and to move at machine speed when that distance collapses.
Where most managed security programs apply generic detection packs to a generic environment model, Cosmos operates from a continuous, bespoke model of your actual environment, your assets, your identities, your cloud topology, your business-critical processes. This matters because the 1.4% of CVEs that get exploited don’t get exploited generically. They get exploited at the specific place where your configuration meets a known adversary technique. Cosmos finds that intersection before an attacker does.
Predictive risk intelligence that is SecOps platform optimized. Rather than processing the full CVE universe at equal weight, Cosmos applies threat-driven prioritization: which vulnerabilities are actively being used against organizations with your platform profile, your industry vertical, and your observable external attack surface? What are the most likely adversary techniques against your specific Splunk ES deployment? Pythia maps OSINT, current APT activity, and your environment model to produce an early warning signal, the exploitable paths most likely to be targeted before they appear in your detection logs. It looks at the current detections inherent in your SIEM/SOAR or SecOps platforms and maps threats against that TDIR footprint.
Custom environment emulation and detection generation. Most MDR providers issue the same detection logic to every customer. TekStream runs adversary simulations through a model of your environment, not a generic lab. Custom detections are generated from those simulations in hours or days, not the two-to-four week approval cycles that define the large-platform approach. When an APT technique evolves to bypass a standard rule set, Cosmos knows your environment well enough to generate a compensating detection within your specific context. When we see a potential threat we automatically generate a search to scan our customers for any indicators in the wild and in parallel, generate the detections needed to prevent that threat in production.
AI-native detection for multi-vector, AI-assisted attacks. Our dynamic AI traps module focused on deception engineering, deploys intelligent, dynamically generated decoys calibrated to look like high-value assets in your environment. Where a human attacker might probe one or two surfaces at a time, an AI-assisted attacker generates simultaneous probe activity across multiple vectors. The AI Trap doesn’t need to match every action to a signature. It detects not only velocity and sophistication, but the reasoning pattern, the multi-vector, low-and-slow reconnaissance behavior that looks like background noise to rule-based detection and looks like an attacker to a trained AI model with full behavioral context.
Human expertise at the point of autonomous action. Every Cosmos autonomous action is governed by TekStream’s operator team. The “expert-on-the-loop” model is not a liability hedge, it’s a strategic differentiator. AI agents in a production environment need a governing intelligence that understands both the threat context and the business context. A containment action that is technically correct but business-catastrophically timed is not a good outcome. Cosmos brings both AI speed and human judgment to every decision that touches your environment.
The Competitive Gap
At Gartner Summit 2026, the competitive picture for large-platform MDR providers was direct. There are a few SecOps providers whose custom detection pipeline runs two to four weeks and routes through their professional services organization. The majority of those providers are laser focused on monitoring and securing the infrastructure that they are responsible for, their endpoints, or network assets, or identity platforms. The threat intelligence signal delivered to customers is a lowest-common-denominator ruleset tuned for the broadest possible install base, not for your specific threat profile.
TekStream’s position is the inverse. We provide a security service, initially focused specifically on predictive AI cyber-defense that builds our services expertise into the platform that supports it, delivered as a managed service, and measured against your specific risk posture. We don’t give you a console. We give you a continuously updated threat model of your environment, proactive detection of the attack paths most likely to be exploited against your specific configuration, and human operators who treat every alert as a business decision, not a ticket.
The vulnerability universe will keep expanding. The OSINT surface will keep broadening. AI-assisted attacks will keep compressing the window between disclosure and exploitation. The organizations that navigate this aren’t the ones that patched the most CVEs. They’re the ones that knew which CVEs mattered in their environment, and had an AI running continuous defense before those CVEs became incidents. We’re committed to getting better and doing it collaboratively with our customers and partners. I don’t know about you, but our workload hasn’t slowed down with AI, we’ve just moved from fastening seatbelts to be safe to solving world peace. Read more about TekStream Proactive Cyber Defense here.
About the Author
Bruce Johnson is Vice President, Solutions at TekStream, where the team builds and delivers TekStream Proactive Cyber Defense, an expert-operated, AI-augmented service that continuously hardens enterprise security operations across existing environments, powered by the Cosmos Cyber Defense intelligence platform. To see the Cosmos data or join the program, reach out!