Why does Synack's vulnerability report only cover 32% of the attack surface?

Synack’s data comes from vetted human researchers testing assets that enterprise customers specifically define and scope for assessment. That scoping means the findings are validated, exploitable, and actionable, but they also represent a subset of the total attack surface. Research from Omdia found that organizations test only 32% of their exposed attack surface on average in a given year, which is the gap Synack’s Sara AI Pentesting is designed to help close.

What does the 2026 vulnerability data tell us about attacker behavior?

Three vulnerability categories grew significantly while nearly all others stayed flat or declined: remote code execution (+39%), brute force attacks (+17%), and content injection (+8%). This pattern suggests attackers are shifting from exploiting exposed surfaces to techniques that bypass the perimeter, gaining access through authentication boundaries, developer toolchains, and edge devices rather than through commonly patched structural vulnerabilities like XSS.

Why can't AI replace human security researchers for penetration testing?

AI excels at automated reconnaissance, known-pattern detection, and continuous coverage at machine speed. But vulnerabilities in dynamic systems, particularly AI/LLM attack surfaces, business logic flaws, and novel exploit chains, require creative adversarial reasoning that current AI models can’t reliably replicate. As demonstrated by SRT researchers in real enterprise engagements, finding vulnerabilities in AI agents required social engineering, escalation reasoning, and adversarial prompting that no automated scanner would have surfaced.

What is the OWASP LLM Top 10 and why does it matter for enterprise security?

The OWASP LLM Top 10 is a framework that categorizes the most critical security risks specific to large language model applications, including prompt injection, insecure output handling, excessive agency, and insecure credential storage. As organizations deploy more AI agents with access to internal systems and databases, these vulnerabilities are increasingly appearing in real enterprise environments. Synack has seen a 120% increase in AI/LLM security missions on its platform as customers work to assess and address them.

What is Sara and how does it work alongside the Synack Red Team?

Sara is Synack’s agentic AI solution, an autonomous red agent that runs continuous attack surface mapping, reconnaissance, known-pattern vulnerability detection, and automated triage at machine speed. Sara operates 24/7/365 and handles the scale and repetitive tasks that benefit from AI, while the Synack Red Team focuses on the depth that requires human expertise: creative exploit chains, business logic attacks, zero-day research, and adversarial reasoning. The two operate as a continuous feedback loop, with Sara handing off an updated attack surface picture so researchers can go deeper from the start.

Jul 8, 2026

The 2026 State of Vulnerabilities: What the Data Misses, According to Our Red Team

Our 2026 State of Vulnerabilities Report surfaces what Synack finds in tested customer environments. At a recent webinar, two of our most decorated researchers from the Synack Red Team describe the threat landscape they’re seeing beyond the report findings. Here's what the data shows, what practitioners have experienced, and what your security program should do about the gap.

Based on Synack’s State of Vulnerabilities report, three vulnerability categories grew in 2025: remote code execution (+39%), brute force attacks (+17%), and content injection (+8%).
Edge devices and developer toolchains have become primary attacker entry points, displacing traditional perimeter attacks.
Time to exploit has compressed from months to hours, and in some cases vulnerabilities are being exploited before they're even published.
AI doesn't replace skilled security researchers; it accelerates what both attackers and defenders can already do.
Synack customers reduced their mean time to remediate by 47% on average.

The data in our 2026 State of Vulnerabilities report tells you what happened inside a defined scope. But keep in mind, attackers don’t have scope.

That was the throughline in our recent webinar with two of our Synack Red Team researchers who provided an on-the-ground perspective of our report. Nicolas Krassas, Head of Threat and Vulnerability Management at Henkel, and Mustafa Can Ipekci, a security researcher with six years of active testing on the Synack platform, spend their days finding the things organizations haven’t thought to look for. Hearing from them proved that it’s worth understanding the gap between what we measure and what attackers actually see.

This post is my attempt to bridge those two perspectives. I’ll walk through what the report shows, what Nicolas and Mustafa are actually encountering in enterprise environments today, and what the combination of those views should mean for your security program.

Why Our Vulnerability Data Is Only Part of the Picture

Our report compiles findings from Synack penetration tests on assets our customers specifically asked us to test. Because we have a strict vulnerability acceptance process, the numbers in our report reflect real, exploitable risk. Every vulnerability goes through a multi-step approval process before it counts as a finding.

But it also means the data has a scope boundary. We reported that cross-site scripting (XSS) was the top vulnerability, which is accurate for what we’re testing, but that’s not true for Nicolas and Mustafa. When our researchers push back and say authorization and access control issues are arguably more prevalent, they’re right too. The difference is that XSS is easier to surface in scoped, unauthenticated testing. Access control issues show up more when you go deeper.

That distinction is why I encourage people to read our report as a directional signal, not an exhaustive inventory. The 2026 State of Vulnerabilities Report tells you what’s surfacing in tested environments. It won’t tell you what’s hiding in the 68% of your attack surface that goes untested.

What the Vulnerability Mix Tells Us

Starting at the macro level: 48,244 CVEs were published in 2025, which represents a 20% year-over-year increase. More vulnerabilities are being disclosed, the threat surface is expanding, and the exploitable findings our researchers are validating grew with it. Synack surfaced more than 11,000 exploitable vulnerabilities across customer environments last year.

Within that, the category mix is what I find most interesting. XSS, which has long dominated our findings, is actually declining year over year. That’s a good sign, suggesting that security programs focused on reducing structural vulnerabilities are working and showing up in the data. But while nearly every other category stayed flat or reduced, three specific categories jumped in 2025:

Remote code execution: up 39%
Brute force attacks: up 17%
Content injection: up 8%

All three are related to gaining access past the perimeter rather than exploiting things already exposed. RCE lets an attacker execute code on your servers. Brute force targets the authentication boundary. Content injection is a vector for credential theft and social engineering. When I see those three moving together, it signals that attackers are adapting. Organizations are doing a better job locking down their exposed surface, so adversaries are pivoting to techniques that bypass the perimeter entirely.

Meanwhile, the exploit timeline, once measured in months, has compressed to hours. We’re even seeing a negative time to exploit, where vulnerabilities are being leveraged before they’ve been publicly disclosed. AI-enabled adversaries running continuous, machine-speed reconnaissance across thousands of assets at once are closing that window faster than any manual security program can respond to it.

What SRT Researchers Are Seeing That the Data Doesn’t Fully Capture

When I asked Nicolas and Mustafa to put the report data in context against what they see in active engagements, their answers sharpened the picture.

The Perimeter Is Shrinking, But the Attack Surface Isn’t

Enterprise environments are closing down their exposed surfaces, to the point where some assessments have narrowed to just a set of firewall appliances because nothing else is exposed. But locking down what’s visible doesn’t make organizations safer. It relocates the risk. Ivanti, Palo Alto, Fortinet and Check Point are now primary attack vectors. Devices that were considered solid perimeter protection for years are routinely compromised. And the risk doesn’t stop at the edge. You can’t test what you don’t know exists, and developer environments are where that blind spot is widest. Internal APIs, staging environments, and other developer assets get spun up and exposed publicly without InfoSec ever being looped in, not because anyone is negligent, but because security and development simply aren’t talking to each other as often as the architecture demands. The result is that the most current, fastest-moving part of the attack surface is also the part security teams have the least visibility into.

The Same Vulnerabilities Have Been Winning for 30 Years

Access control and authorization failures are still the most prevalent issues in enterprise environments. They just cycle through new technology stacks. Every time a new AI model ships, the assumption is that the coding problems are finally solved. They aren’t. XSS gets the attention because it’s easy to surface in scoped, unauthenticated testing, but authorization issues are what’s actually showing up most in active enterprise environments. AI hasn’t fixed that. It’s given developers a faster way to ship the same mistakes at greater scale.

Defenders Have Scope. Adversaries Don’t.

When a client tells us not to test brute force because they don’t want accounts locked, that’s a reasonable constraint for a managed engagement. Adversaries don’t honor that constraint. They test everything, including the attack surface that most organizations never get to. The coverage gap isn’t just a blind spot. It’s a target list. Adversaries will find what’s in that untested portion. The question is whether you find it first.

What AI Attack Surfaces Are Revealing About Human Creativity

In our webinar, I also brought up how we’ve seen a 120% increase in AI and LLM security missions on the Synack PTaaS platform. The findings from those engagements point to two consistent problems that no scanner is going to catch.

AI Agents With Privileged Access Are a New, Largely Untested Attack Surface

Developers are deploying AI agents with access to systems and data they were never designed to interact with. Mustafa described one engagement where he essentially tricked a developer-facing AI agent with direct database access. He asked it to query a table “just to confirm it wasn’t hallucinating.” That ask escalated to a direct SQL Server callback. As a low-privileged user, the result was full control over the DBMS and access to 65 databases. The agent was deployed in a UAT environment, but the path to production was clear.

In another engagement, Nicolas was testing a voice-chat application that used two LLMs in sequence, with a guardrail between them. The credentials for the second model were stored in accessible files. Nicolas bypassed the guardrail entirely, making direct API calls with no restrictions, and wrote a Python program to interact with the model freely, the same way you’d query any GPT model.

This Attack Surface Can’t Be Evaluated With Automated Tooling

Both examples map to the OWASP LLM Top 10, which includes excessive agency, insecure credential storage, and inadequate guardrails. Both required social engineering, adversarial prompting, and the creativity to probe for unintended behavior in a system that was behaving exactly as designed, just in ways its developers hadn’t anticipated. That’s the core problem with claims that fully autonomous AI can adequately test AI attack surfaces. Dynamic systems require dynamic attackers. That still means humans.

How Synack Customers Are Responding

Despite these firsthand stories, there are also encouraging data points in the report: Synack customers reduced mean time to remediate vulnerabilities by 47% on average. Critical vulnerabilities dropped from 63 days to 38. High vulnerabilities nearly cut in half. That reflects real progress: better visibility into exploitable findings, tighter integration between security and development, and more actionable reporting that developers can actually use.

But faster remediation only applies to vulnerabilities you’ve already found. The penetration testing coverage gap means that if you’re improving metrics on the 32% of your attack surface being tested while the other 68% remains invisible, the progress is real but incomplete. The benchmark for “fast enough” has also shifted. Exploits that used to take months to materialize are now appearing in hours. The improvement curve is real, and the expectations curve is accelerating faster.

The model we’re working toward at Synack is designed to address both sides of that equation. Sara, our agentic AI solution, handles attack surface mapping, continuous reconnaissance, known-pattern vulnerability detection, and automated triage at machine speed, 24/7/365. That frees up SRT researchers like Nicolas and Mustafa to focus on the things machines can’t do: creative exploit chaining, business logic attacks, novel zero-day research, and the adversarial reasoning that no scanner replicates. Sara handles the scale. The SRT handles the depth. Together, they are the engine behind continuous security validation at scale, so researchers aren’t starting from scratch and organizations aren’t waiting for the next annual test to find out what changed.

How to Close the Penetration Testing Coverage Gap

If any of this maps to what you’re dealing with in your security program, there are two practical next steps.

First, check out the 2026 State of Vulnerabilities Report. It’s the most complete view we have of what’s surfacing in tested environments right now, and it’s useful context for benchmarking your own vulnerability mix against what we’re seeing across the industry.

Second, if the 68% coverage gap is the problem you need to solve, start a Sara AI Pentest free trial. It deploys in days, covers your actual attack surface with agentic AI, and delivers validated findings with no long-term commitment. You’ll see what’s in your untested environment, and that visibility is the first step to addressing it.