Vulnerability assessment by AI alone drops from 29% to 9%. Why the illusion that 'leaving it to AI is fine' has collapsed
Until last year, there was a sentiment that 'we should just leave it to AI.' In the world of security, it was the same.
Vulnerability scanning using AI. It automatically searches for holes in code and reports them automatically. AI was supposed to take over the tedious work that humans used to do... that was the idea.
But when the 2026 data came out, the story changed completely.
Companies that 'perform vulnerability assessments using only AI' plummeted from 29% to 9% in just one year. That means not even one in three companies trusts it anymore. What happened?
The numbers tell the story of the 'collapse of trust in AI'
Cobalt, a provider of penetration testing platforms, released its 2026 'State of Pentesting Report.' The subjects were about 450 cybersecurity experts. It is a survey that tracked changes over one year by asking the same questions in 2025 and 2026.
Looking at the numbers might be a bit surprising.
Companies performing vulnerability tests using only AI: 29% → 9% (20 percentage point decrease)
Companies that prefer a hybrid of human and AI: approximately 47% (22 percentage point increase year-over-year)
Companies that responded that fully automated AI scanning missed critical vulnerabilities: 78%
From 'trusting AI completely' to 'relying too much on AI is dangerous.' I thought, it's amazing how much things can change in just one year.
Why does AI 'miss' things?
The essence of the problem lies in the difference between what AI can find and what it cannot.
AI scanning tools are good at so-called 'known patterns.' Vulnerabilities that have been turned into databases, such as SQL injection and cross-site scripting (XSS), where it is known that 'if this type of attack comes, it is dangerous.' They are fast and accurate here.
But vulnerabilities specific to AI apps are different.
For example, 'prompt injection.' This is an attack that causes an app equipped with an LLM (Large Language Model) to perform unintended actions through clever manipulation of language. To prevent this, you cannot judge where the danger lies unless you deeply understand the design and context of that app. Just throwing single automated queries at it will never find it.
Cobalt's CISO (Chief Information Security Officer) says, 'LLM vulnerabilities are invisible to tools that do not understand the application's architecture.'
In other words, AI can find 'stereotypical weaknesses,' but it doesn't understand 'contextual or design issues.'
Vulnerabilities in AI apps are 2.7 times more 'dangerous' than normal ones
There is one more piece of data that cannot be overlooked.
According to Cobalt's data, among the issues detected in penetration tests of AI applications incorporating LLMs, the proportion classified as high-risk is 2.7 times that of traditional software. While the overall high-risk rate is about 12%, about 32% of AI-based applications emerge as high-risk.
Furthermore, the story that follows is terrible. Of the vulnerabilities found, only 38% were fixed. The remaining 62% are left 'unaddressed.' This is the lowest fix rate in all categories.
The time taken for fixes also doubled from 19 days to 36 days. This doesn't simply mean 'responses have become slower,' but also that 'we have started dealing with more complex and difficult problems.' It requires coordination with model vendors or fundamental changes to the architecture. Such responses take many times longer than applying a patch.
The percentage of security experts who feel they are 'keeping up with AI security' dropped from 64% in 2025 to 51% in 2026. The number of people who lack confidence is increasing, isn't it?
A 'new threat list' when using AI
As more companies use AI, attacks will start coming through dedicated entry points.
Looking at data from companies that have actually experienced AI-related security incidents, the most common causes are as follows.
Shadow AI (44%): Starting to use AI tools not permitted by the company, leading to information leaks
Data or model poisoning (41%): Tampering with the data AI uses for learning or the model itself
Inappropriate output handling (41%): Using AI output as-is, causing unintended processes to run
Supply chain vulnerabilities (35%): The AI APIs or libraries being used are exposed to danger
Prompt injection (34%): Manipulating language to make AI perform unintended actions
It's somewhat understandable that Shadow AI is the most common, isn't it?
Thinking 'ChatGPT is convenient,' people paste work information into personal accounts to process it. That becomes a corporate risk just like that. It's the reality that the speed of tool adoption far exceeds the speed of establishing security policies.
Not 'AI or human,' but 'where to use which'
So, should we stop using all AI security tools? Of course not.
The current trend is 'hybrid'.
Leave the parts that cover a wide range quickly to AI. However, for diagnosing important systems and AI applications, always include visual inspection and judgment by human experts. Designs are becoming like that.
What Cobalt points out is that 'the problem isn't automation versus manual, but the design of which judgments to entrust to AI is important.' Separate the parts that should be automated from the parts that humans should look at. Whether this distinction is made determines the difference in security levels.
What I thought was interesting is that it can be read as 'humans have started tackling more serious problems.' The fact that it takes longer to fix might not be a step backward, but rather that we have finally started working on the truly difficult problems that had been put off until now.
📌 Miccell's Perspective: The AI boom has been pulled down by the 'gravity of practical work'
What left the biggest impression on me after reading this report wasn't the expression 'collapse of trust,' but rather that the perspective of converging toward appropriate usage felt more accurate.
At the beginning of an AI boom, there's an atmosphere in every tool that says, 'Let's leave everything to AI.' It's a period where the dream of efficiency runs ahead of reality. But as you actually use it, you start to see areas where 'AI just can't handle this.' And then the design changes.
In the world of security, that 'gravity of practical work' is particularly strong. If there's a hole, you really will be attacked. It's a world where you can't lie, because 'I thought AI was doing a good job' isn't an acceptable excuse.
In the stage where AI 'establishes itself as a tool,' this kind of backlash is inevitable.
I think the data from the security industry this time can be read as a trailer for the fact that the same thing might happen later in other fields. I believe that people and organizations who quickly move past the 'period of blind faith in AI' and shift to 'designing for smart AI use' will be the ones who become stronger in the next phase.
🔚 Summary: Toward an era of 'designing AI' rather than 'trusting AI'
Companies trusting vulnerability scanning by AI alone: Plunged from 29% to 9% in just one year
78% have experienced 'misses' by AI automated scanning
AI apps carry 2.7 times more high-risk issues than traditional software
A 'mountain of difficult problems' with a 38% fix rate and 36-day fix time
The answer isn't 'AI vs. Humans,' but in the design of roles
Using AI for security itself isn't bad. But the expectation that 'AI will do everything' is quietly being refuted by this year's data.
It might be worth taking a moment to look back at the 'parts you're leaving entirely to AI' in your own organization or work.
For any tool, whether it becomes a game-changer or a liability depends on the design. I felt that the security world taught us that through data.
🛠️ Production Notes
At first, this theme looked like a negative story about 'AI vulnerability scanning tools losing trust.' But as I followed the numbers, I began to read it as a 'perfectly natural cycle where feedback from practical work optimizes design.'
The feeling of being between the 'illusion phase' and the 'adaptation phase' of the AI boom might apply not just to security, but to all the tools we use.
The fact that LLM vulnerabilities are the 'least fixed category' is frankly scary data. The numbers convey that security departments at companies using AI products are currently struggling with 'design-level problems.'
Miccell - Once you understand how things work, the world becomes much more interesting.