When AI Starts Doing the Hacker's Work
A recent cybersecurity research project has highlighted something that's becoming harder to ignore: AI is getting very good at helping with the kind of technical work that used to require years of experience, specialized skills, and a lot of time.
Researchers at Hacktron AI reportedly managed to exploit a vulnerability affecting an image-processing library used by OpenAI's community platform. What makes the story especially interesting is that they used Anthropic's Claude Opus 5 to help them along the way.
And according to the researchers, it made a real difference.
It Started With a Pretty Ordinary Vulnerability
At the heart of the incident was a heap overflow, a type of memory-corruption bug that can sometimes allow an attacker to take control of how a program runs.
The vulnerability was found in libheif, a library used to process images. It was being used by Discourse, the platform behind OpenAI's community forum.
The researchers then found that the vulnerability could be combined with a weakness in the company's single sign-on (SSO) setup.
That combination was the real problem.
Something as harmless-looking as uploading an image could potentially be turned into remote code execution (RCE), giving an attacker a way to run code on the target system.
This Is Where AI Changed Things
Finding a vulnerability is one thing. Turning it into a working exploit is another.
According to the researchers, they initially struggled to automate the process using an earlier version of Claude. The breakthrough reportedly came after Anthropic released Opus 5.
The model helped them work through some of the technical challenges involved in bypassing memory protections such as ASLR. The researchers say they were then able to complete the exploit in just a few hours.
That's what makes this story worth paying attention to.
AI isn't just being used to write code or find bugs anymore. It can also help security researchers work through complicated offensive-security problems that would traditionally require a lot of specialized knowledge.
The Surprising Part? The Cost
There's another number that stands out: less than $3,000.
Hacktron AI says its broader research campaign, called "HEIF Heist," cost less than $3,000 in API usage.
That doesn't mean a real-world cyberattack against a major company would cost only $3,000. There's a lot more involved in an actual operation, including infrastructure, reconnaissance, human expertise, and maintaining access.
Still, the figure is hard to ignore.
Work that once demanded expensive expertise and weeks of engineering can potentially be compressed into a few days — or even hours — when AI is added to the equation.
Why Companies Should Care
For years, sophisticated memory-corruption attacks were difficult for a simple reason: they were difficult.
They required deep technical knowledge, patience, and a lot of trial and error.
AI changes that.
A capable model can analyze code, suggest possible attack paths, help troubleshoot failed attempts, and automate parts of the process. That doesn't necessarily mean AI can carry out an entire attack on its own. But it can make a skilled person significantly faster.
And that's probably the most important takeaway here.
Security teams have traditionally asked:
Can this vulnerability be exploited?
Increasingly, they also need to think about:
How quickly could someone exploit it with AI helping them?
OpenAI Was Told About the Vulnerability
The researchers reportedly disclosed their findings to OpenAI.
The company thanked them for getting in touch and sharing what they had discovered.
There's also an important distinction to make here. This was reported as a security research exercise, not a criminal attack carried out against OpenAI.
Showing that a system can be compromised under controlled research conditions doesn't automatically mean that criminals used the same technique in the wild.
AI Is Becoming Part of the Cybersecurity Arms Race
The bigger story isn't really about OpenAI or Anthropic.
It's about what happens when the same technology can be used on both sides of a security battle.
Defenders can use AI to find vulnerabilities, review code, monitor systems, investigate suspicious activity, and respond to incidents.
Attackers can potentially use it to analyze software, automate reconnaissance, develop exploits, and work through technical roadblocks faster.
In other words, we're heading toward a world where automation is fighting automation.
That doesn't make the basics of security any less important. Keeping software updated, securing authentication systems, limiting access, isolating sensitive services, and monitoring unusual activity are still essential.
But there's a new variable to consider: speed.
If AI can help someone move from discovering a vulnerability to exploiting it much faster, organizations have less time to react.
The Bigger Picture
The "HEIF Heist" research offers a glimpse of where cybersecurity may be heading.
The interesting part isn't simply that an AI model helped researchers exploit a vulnerability. It's how much time AI may be able to save when the work involves complex technical problems.
Tasks that once took days or weeks of experimentation can potentially be shortened dramatically.
For security teams, that's a major shift.
AI is becoming both a security tool and part of the threat itself.
And the race between offensive and defensive AI is only getting started.