You have 1 article left to read this month before you need to register a free LeadDev.com account.
Estimated reading time: 4 minutes
Autonomous AI agents are showing early promise in spotting software flaws before attackers do, potentially reshaping the cybersecurity landscape.
The rise of artificial intelligence has long been viewed as a double-edged sword in cybersecurity, amplifying threats in the hands of attackers, while offering promise for defense.
Now, rather than simply detecting known vulnerabilities, a new wave of AI agents have the potential to proactively identify unknown flaws, simulate attacker behavior, and even suggest or implement fixes. Although still in its early stages, this technology has the potential to reshape the future of vulnerability management and shrink the window of opportunity for attackers.
AI agents in action
Your inbox, upgraded.
Receive weekly engineering insights to level up your leadership approach.
In one recent case, Google revealed that its in-house AI agent, Big Sleep, helped detect and prevent a real-world exploit, marking what the company believes is the first time an AI agent has done so in the wild.
It’s a milestone that hints at the broader potential of this technology. Unlike conventional scanners, AI agents can reason through complex systems, adapt to new environments, and uncover subtle flaws, not just in application code, but across entire software stacks, including cloud configurations, access controls, APIs, and even business logic, with minimal human input.
Another promising example is Code Intelligence’s Spark, an LLM-guided fuzzing agent. Traditional fuzzing tools work by blindly throwing random inputs at software to identify what breaks, but Spark takes this a step further by utilizing an LLM to understand the code’s intent and generate more intelligent, targeted inputs that are likely to expose hidden bugs, crashes, or security vulnerabilities.
In 2024, Spark discovered a serious heap-based use-after-free vulnerability in the wolfSSL library. It has since flagged multiple real-world bugs across popular open-source projects – including a dynamic stack buffer overflow in Google’s Abseil C++ library – positioning itself as more of a self-directed auditor than a passive tool.
Meanwhile, researchers at UC Berkeley have taken a multi-agent approach. As part of a 2024 research project under the so-called CyberGym initiative, teams deployed a swarm of AI agents across nearly 200 open source codebases. These agents, which were based on multiple AI toolkits including Cybench, ENiGMA and OpenAI’s Codex, operated with distinct roles: one acted as a code reviewer, another as an attacker, and another as a tester, collaborating to discover 17 bugs, including 15 previously unknown zero-day vulnerabilities. Each agent worked independently but shared findings to develop proof-of-concept exploits, demonstrating a level of coordination rarely seen in automated vulnerability research.
More like this
Not without caveats
Despite these advances, experts caution that AI agents aren’t yet ready to take the wheel alone. “While AI agents have the potential to improve cybersecurity, they’re still not at a point where they can replace human expertise,” Nicolette Carklin, technical specialist at SecureFlag, told LeadDev. “Many vulnerabilities are subtle or context-dependent, and current models can easily miss them. These agents also need access to sensitive systems – and if misconfigured or compromised, they could introduce new vulnerabilities instead of preventing them.”
There’s also the risk of overtrusting what the AI tells you. Just because a model flags something doesn’t mean it’s right, or even relevant. As Holly Foxcroft, cybersecurity business partner at OneAdvanced, puts it: “AI still needs human oversight to ensure accuracy, prioritization, and context, especially where business operations and customer trust are on the line. The risks of overreliance include hallucinated threats or missed context that create a false sense of security.”
That said, Foxcroft is optimistic about the bigger picture: “From a business standpoint, this isn’t just a technical win, it’s a strategic one,” she said. “The ability to detect issues earlier can help reduce the cost and disruption of late-stage fixes, regulatory exposure, and reputational harm.” The key, she says, is to treat AI agents as a complement to, not a replacement for, secure development practices and existing security controls.
Still, the rise of autonomous agents raises deeper questions about how security teams might evolve. If agents can autonomously scan, reason through, and even patch software at scale, what does that mean for headcount, skillsets, or hiring? While few expect mass layoffs, some believe we’re heading toward leaner, more orchestrated teams where human analysts focus less on writing detection rules and more on managing fleets of agents and interpreting their output.
The UC Berkeley CyberGym project offers a glimpse into that future. Researchers didn’t just test single agents in isolation; they coordinated swarms of agents, each with different capabilities and reasoning styles, to compete or collaborate on vulnerability detection. In that world, the ability to orchestrate multi-agent workflows could soon become one of the most valuable skills in cybersecurity.
The road ahead
The current generation of AI agents show real promise, and the technology is improving every day. But widespread adoption will require overcoming several hurdles. This includes technical maturity, in the sense that agents must become more reliable, secure, and scalable before they can be trusted in real-world environments; trust in the results, and seamless integration into existing workflows.
While several major vendors, including Google, Microsoft, and Palo Alto Networks, are now rolling out AI agent capabilities within their security platforms, organizations also need to invest in training and policy frameworks to ensure these tools are deployed effectively and responsibly.
Still, the direction is clear. Where traditional scanners and manual testing methods often fall short, bogged down by scale, complexity, and the limits of human attention, AI agents can bring speed, breadth, and persistence. With the right guardrails, they could become powerful allies in the race to find flaws before adversaries do.
As these tools mature, they may not only detect vulnerabilities but also help prioritize, explain, and even autonomously remediate them, augmenting human teams rather than replacing them. The future of vulnerability management won’t be AI alone, but rather AI in lockstep with people, processes, and policies.
