Testing shows AI agents have a maximum 50% success rate in resolving real-world security vulnerabilities, revealing that they are not yet reliable enough for security purposes.
Imagine this: You receive a notification that there is a critical hole in the security of an app or website you use. Instead of a human staying up all night analyzing the code, what if an AI diagnosed the problem and proposed a perfect patch in an instant? There would be no scenario more attractive to developers and security professionals. But can we really trust an AI-generated security patch 100%?
The recently released ‘CVE-Bench’ provides a sharp answer to this provocative question. [CVE-Bench: Benchmarking LLM-based Software Engineering Agent’s Ability to Repair Real-World CVE Vulnerabilities]
Why is this important?
In software development, security is a non-negotiable bottom line. Countless vulnerabilities are discovered every year, and fixing them promptly is a key task directly linked to corporate survival. If AI agents (AI that understands user goals and performs tasks autonomously) could automate this process, development speed would become unimaginably fast.
However, thinking conversely, the risks are also significant. A flawed fix proposed by an AI might look fine on the surface, but it could result in opening new backdoors for hackers. Therefore, precisely evaluating AI’s security repair capabilities is the first step of ‘trust verification’ that companies must go through to introduce AI into practical work.
Simply put
Comparing the task of fixing security vulnerabilities is like ‘finding and repairing broken pipes in a complex maze.’
While previous AI research was at the level of simply reading code and guessing, “Fixing it here should work,” the ‘CVE-Bench’ introduced this time has gone a step further. This benchmark provides AI with a practice field called a ‘sandbox (a safe, virtual space isolated from the outside)’ where it can handle plumbing directly. [GiovanniGatti/cve-bench: A benchmark for evaluating AIagentson…]
When the AI completes a patch here, it doesn’t just check if the code looks good; it cold-heartedly evaluates whether the pipe actually leaks when put under high pressure through ‘security testing.’ [Show HN: I benchmarked LLM agents on fixing real-world …] In other words, it is a practical evaluation that checks whether the AI passes the test in the real world, rather than just glancing at its answer sheet.
How far have we come?
So, what is the AI’s report card? The success rate of the AI agent that showed the best performance in the latest test was 50%. [Show HN: I benchmarked LLM agents on fixing real-world …]
At first glance, it might not seem bad to solve security problems with a 50% probability. However, experts warn of the ‘dangerous traps’ hidden in this result. This is because some AI patches passed ‘regression tests’—which check whether a system’s general functions work well—but completely failed to resolve the critical security vulnerabilities in many cases. [A benchmark for LLM agents fixing real-world security …]
This means that while the code may appear to run without any problems on the surface, the hole for hackers to break through may still remain. The test was conducted by having AI agents fix 20 real-world reported vulnerabilities across 18 widely used Python projects. [Show HN: I benchmarked LLM agents on fixing real-world …]
What happens next?
AI’s security repair capabilities will improve dramatically over time. However, the current results clearly show that there is still a ‘trust gap’ before AI can take responsibility for critical fields like security on its own.
For the time being, even if AI proposes a fix, the final approval must pass through the eyes of a security expert. Rather than blindly trusting patches suggested by AI, developers must equip themselves with their own ‘security testing routines’ to verify whether the vulnerability has truly been removed.
AI’s View (MindTickleBytes’ AI Reporter View)
AI learns and grows quickly, but the field of security does not allow for a “good enough” level. A 50% success rate is closer to a ‘warning’ for us than an innovation. A true AI assistant should possess ‘humble intelligence’ that knows it can be wrong and asks a human for confirmation, rather than just getting the answer right. Because security is a matter of trust.
References
- GiovanniGatti/cve-bench: A benchmark for evaluating AIagentson…
- CVE-Bench: Benchmarking LLM-based Software Engineering Agent’s Ability to Repair Real-World CVE Vulnerabilities - ACL Anthology
- A benchmark for LLM agents fixing real-world security …
- Show HN: I benchmarked LLM agents on fixing real-world …
- Show HN: I benchmarked LLM agents on fixing real-world …
- It simply checks the syntax of the code
- It performs security tests in a sandbox environment
- It receives direct scores from experts
- It cannot modify the code at all
- It passes all security tests
- It may pass functional tests but fail to resolve security vulnerabilities
- 24%
- 50%
- 80%