The
AI arms race in cybersecurity is continuing to accelerate, with bad actors adopting the technology to enhance and automate their campaigns while security pros increasingly match those efforts, using the emerging technology to push back against attacks while saving time on the backend.
That is the finding in
HackerOne’s
9th Annual Hacker-Powered Security Report, released this week. In the report, the crowd-source security platform provider noted a year-over-year surge in the number of prompt injection exploits and valid AI-related vulnerabilities, and a rapid increase in organizations adopting AI and in security researchers using AI tools in their workflows.
Researchers’ growing embrace of the technology is a matter of both necessity and maturity, according to
Naz Bozdemir, HackerOne’s lead product researcher.
“Defenders know attackers are experimenting with AI, from writing more convincing phishing campaigns to automating exploit chains, and they need to keep pace,” Bozdemir told MSSP Alert. “At the same time, comfort with AI is growing because the benefits are real: researchers report that AI saves them hours each week on reconnaissance, payload generation, and report writing, freeing them to focus on harder, higher-impact flaws.”
Here Comes the 'Bionic Hacker'
In its 31-page report, San Francisco-based HackerOne focused on the rise of what it calls the “Bionic Hacker,” - security pros who are using AI to enhance their hunting and other capabilities, such as pattern recognition, security issue discovery, and reconnaissance.
The numbers in the report – culled from surveys of customers and researchers and the company’s vulnerability database – illustrate the buildup of AI capabilities by defenders and attackers.
Program adoption among organizations expanded by 270% – 1,121 distinct customer programs included AI in scope – and 70% of researchers now use AI tools in their workflows, giving rise to the Bionic Hacker. In addition, 58% of survey respondents said they’re improving their skills in AI and machine learning security, and 41% are already testing AI assets as part of their regular work.
Punch, Counterpunch
All of that will be important as threat actors work just as hard to incorporate AI in their nefarious activities. Prompt injection vulnerabilities – which allow hackers to put malicious instructions in prompts for AI systems to manipulate them into running unintended actions and bypassing guardrails – jumped in number by 540% year-over-year, marking them as the top AI security threat.
In addition, there was a 210% rise in valid AI-related vulnerabilities and a 335% jump in the number of bounties paid for AI vulnerabilities.
“The cat-and-mouse game hasn’t ended; it’s intensified,” Bozdemir said. “AI creates a symbiotic advantage for whoever uses it. AI drives speed and scale, while humans deliver the business context and expertise that turn findings into action.”
An Intensifying Environment
According to the report, Model Context Protocol (
MCP) servers – which enable large language models (LLMs) to extend their reach to external data and tools – are helping to drive security researchers’ use of AI, with 21% of researchers using HackerOne tools developing MCP-specific security tools. Meanwhile, 82% of such researchers are developing skills to fight against prompt injection attempts.
That intensity in the AI-vs.-AI security space will ramp up as
adoption of AI agents does. Attackers are scaling their use of “hackbots,” a type of AI agent that, between July 1, 2024, and June 30, submitted more than 560 vulnerability reports on HackerOne’s platform, almost half of which were valid, Bozdemir said.
“Their strength lies in high-volume, commodity issues, but they still fall short on the hardest problems: business logic errors, multi-step exploits, and privilege escalations remain the domain of human ingenuity,” she said.
Agents for Security Pros
For defenders, AI agents are reshaping workflows and automating reconnaissance, triage, and prioritization, saving the equivalent of an extra week of analyst time each month, Bozdemir said.
“Just as importantly, defenders report that AI improves focus, streamlining operations and freeing up time for humans to focus on systemic design flaws, chained exploits, and authorization bypasses that machines can’t address alone,” she said.
HackerOne is also previewing Hai, its agentic AI system aimed at streamlining communications, improving report quality, and accelerating impact, with three out of five users pointing to time savings and efficiencies as the largest impact. Of those surveyed, 60% said Hai saved four to eight hours of work per month, while 20% said the time savings were 24 to 40 hours a month.
In all, 90% of the vendors' customers have adopted Hai, according to HackerOne.
MSSPs: Don't Forget the Human Touch
Given their growing central role as security providers for companies – including many SMBs – the information in HackerOne’s report should resonate with MSSPs and MSPs.
“Their adoption of AI directly determines the resilience of thousands of downstream enterprises,” Bozdemir said. “The challenge is that attackers aren’t just using AI for speed, they’re using it to generate phishing campaigns, create deepfakes, and automate exploitation chains.”
The numbers from the report underscore how quickly the threat surface is expanding, she said. However, the automation alone that AI brings won’t solve the entire problem. AI tools can do some tasks well – reconnaissance and triage being two – but they can miss complex flaws, business logic errors, multi-step exploits, and other security situations. MSSPs that rely only on automation, without human augmentation, will expose their clients to more threats.
“What works is a layered approach,” she said. “MSSPs should combine AI-enabled triage for speed with continuous adversarial testing powered by a global researcher community for depth. The data shows enterprises that layered bug bounties, pentests, and disclosure programs with AI red teaming saw 65% stronger outcomes.”
