Vibe Hacking Not Yet Possible

Security experts are closely tracking the event horizon when artificial intelligence tools become good enough for anyone to launch mass hack attacks with minimal effort. Today is not that day.

See Also: On Demand | Global Incident Response Report 2025

That assessment comes from researchers at Forescout who tested how 50 large language models performed in simulations of real-world attacks. They found that none can yet track the full course of vulnerability identification to exploit development.

Code-illiterate but AI-enabled script kiddies able to wreak havoc by weaponizing software vulnerabilities into automated exploits remains but a future possibility. So-called “vibe hacking” is not yet possible.

The term is a variation on vibe coding, which is jargon referring to having AI write usable code, even if the user has no idea how or why the code works, and they’re happy to work around or overlook AI-generated bugs or glitches along the way.

Researchers evaluated three different types of LLMs: open-source models hosted on HuggingFace, underground models such as WormGPT, EvilAI and GhostGPT – all available via cybercrime forums or Telegram channels – and commercial models such as OpenAI’s ChatGPT, Google’s Gemini, Microsoft’s Copilot and Anthropic’s Claude. The research ran from February to April.

Researchers subjected each LLM to two types of vulnerability research tasks: one a simple task to establish a baseline, and another more complex job. They found 48% failed the first task, and 55% the second. They instructed the successful models to develop an exploit for each vulnerability. Their failure rates respectively jumped to 66% and 93%.

Overall, the researchers found massive variability based on the type of LLM used:

• Open-source models: The 16 tested were “unsuitable even for basic vulnerability research”;
• Underground models: The 23 tested were “hampered by usability issues, including limited access, unstable behavior, poor output formatting and restricted context length”;
• Commercial models: The 18 tested were often restricted by guardrails, and “only three models succeeded in producing a working exploit,” and then only with extensive guidance by expert users.

To the last point, the researchers didn’t approach these tests as if a newbie was querying LLMs. Rather, they pretended to be an experienced security researcher using the LLM to assist in their investigation of a vulnerability. Then they pretended to be an experienced penetration tester who was using an LLM to help them develop an exploit. “These were interactive prompts requiring collaboration to build, test and debug the exploit,” the researchers said.

None of the LLMs succeeded by themselves. “No single model completed all tasks, underscoring that attackers still cannot rely on one tool to cover the full exploitation pipeline,” the report says.

No Vibe-Hacking Event Horizon

The extensive amount of user guidance required to make the commercial models deliver a working exploit, when they could, belies how such technology remains ill-suited for novices, at least when attempting to research flaws and develop exploits.

Even when an LLM got something wrong, it still sounded right. “The confident tone of LLM-generated responses, even when incorrect, can mislead inexperienced attackers, ironically the group most likely to rely on them it,” the report says.

The term “vibe coding” may have originated with San Francisco resident Andrej Karpathy, a co-founder of OpenAI, who said in a February post to social platform X: “There’s a new kind of coding I call ‘vibe coding,’ where you fully give in to the vibes, embrace exponentials and forget that the code even exists.”

He likened LLM assistance to being on a slider, with traditional programming circa three years ago all the way to the left and vibe coding being all the way to the right.

Despite its potential, moving the slider all the way to the right still doesn’t deliver fully automated vibe coding. “I’m still doing way too much,” Karpathy said.

The ability to do vibe coding does continue to improve. As it does, so too will the ability to use these capabilities for malicious purposes.

“We’re going to see vibe hacking. And people without previous knowledge or deep knowledge will be able to tell AI what it wants to create and be able to go ahead and get that problem solved,” Katie Moussouris, the founder and CEO of Luta Security, recently told Wired.

So far, no security expert reports seeing anything akin to vibe hacking in the wild.

“Threat actors are using these tools to help them with scripting, to help them write spear-phishing emails so that they look authentic, so we know that they’re using them for sort of these normal functions,” Sandra Joyce, vice president of Google Threat Intelligence, told Information Security Media Group at the RSAC Conference in May (see: Breaking Through the Hype of AI in Cyber).

Across the more than 1,300 cybersecurity incidents to which Google’s Mandiant incident response teams responded over the preceding 12 months, “what we have not seen is any real game-changing activity yet,” she said. “We still haven’t seen an incident or breach of a network that has been done with AI in any way that couldn’t have been done by a normal human task.”

Incident responders nonetheless watch closely for when that happens.

AI tools continue to rapidly improve, as Forescout’s researchers reporting seeing just over the course of their three-month research period. Models fine-tuned for certain types of activities, in general, also consistently performed better, as did newer models, which points to how future breakthroughs might happen. “Emerging ‘agentic AI’ models, capable of chaining actions and tools, could further reduce the user burden, especially in exploit development scenarios that require debugging, tool orchestration and feedback loops,” said Michele Campobasso, the researcher at Forescout Vedere Labs who authored the company’s new report.