The Pulse
AI Shrinks the Time Needed to Find and Exploit Vulnerabilities
A CTech survey of 30 security researchers finds that commercial AI tools are shortening the path from vulnerability discovery to working exploits. Researchers also warn that AI agents, open-source infrastructure and automated security opera
AI.info Team ·
Nearly one-third of the security companies surveyed by CTech say commercial AI tools are compressing the time required to discover and develop exploits, turning work that once took months into a process that can take days. The findings come from interviews with 30 security researchers working across Israel’s cyber sector and point to a narrower gap between finding a flaw and using it.
Researchers describe the change as a shift in both sides of the security contest. AI helps defenders examine code, generate proof-of-concept attacks and search large attack surfaces, but the same tools can give smaller teams access to capabilities that once required major funding, specialist expertise or state-level resources. CTech’s survey also identifies risks tied to AI agents, third-party software and the cost of automated security analysis.
One researcher, one day, one commercial model
Avital Leshem, Group Manager of Exposure Management at Check Point, says attackers can now automate attacks within minutes that previously demanded resources available only to highly sophisticated organizations. Her warning frames the central concern raised by the survey: the time available to identify, fix and verify a vulnerability may be shrinking faster than security teams can adjust.
Shoham Danino, Senior Researcher at Surf AI, says implementation was historically the main bottleneck. Turning a theoretical weakness into a working proof of concept could take months; with AI assistance, his team can do it “in a matter of days.”
Five companies surveyed by CTech described small teams or individual researchers finding zero-day flaws in widely used commercial software, including Zoom, Google Cloud and Microsoft Copilot. Alex Mor, Security Research Manager at A Security, said a team member found a zero-day in Zoom in one day using commercially available AI models. “Building an attack like that against closed commercial software used to take a government, a team of experts, and months of work,” Mor said.
AI finds the easy flaws first
Researchers do not describe AI as a replacement for experienced analysts. One-fifth of the companies surveyed said the technology is effective at detecting surface-level weaknesses but often loses context, generates false findings or pursues an incorrect hypothesis for too long.
Idan Revivo, Head of Security Research at Island, summarized the trade-off bluntly: “AI made it cheap to produce an answer and expensive to trust one.” Uri Katz, Director of Research at Oligo Security, compared the moment to the spread of automated fuzzing, which made it easier to find basic bugs before researchers moved on to harder problems.
“Right now you can throw AI at a project and it'll find the low-hanging fruit,” Katz said. He expects that basic attack surface to become increasingly well covered, leaving human researchers to investigate more complex interactions between software, permissions and network controls.
Agents create a new route around credentials
The survey finds that enterprise risk is shifting beyond stolen usernames and passwords. One-fifth of respondents said attackers are increasingly targeting the permissions granted to trusted AI agents and attempting to influence the material those systems consume.
“The threat AI creates is not about what AI models say, but about what agents do,” said Tamir Ishay Sharbat, Director of Security Research at Zenity. His team demonstrated an indirect prompt-injection attack affecting Microsoft Copilot at Black Hat USA 2024. In the scenario described by Sharbat, an attacker could use an email to manipulate the assistant, mislead a victim, conduct phishing and obtain sensitive information.
Zenity was one of four companies in the survey to publish research showing zero-click data exfiltration or takeover through AI agents consuming shared documents, emails or web pages. Omer Nissim, a Security Researcher at Sweet Security, described a separate case in which an internal AI assistant disclosed information that had protected an environment from outside access.
Security automation has a price tag
AI is also giving attackers a way to target the operating costs of security teams. Security operations centers increasingly use models to read, enrich and reason about alerts before a human reviews them, with each step consuming tokens.
Three respondents identified a form of economic attack in which adversaries flood automated security systems with large volumes of meaningless data. Ido Shtrauch, Security Research Team Lead at Conifers AI, said attackers can force companies to spend their AI budgets processing noise.
“Much of the volume was not hiding anything at all,” Shtrauch said. “Analyzing it costs us every single time. Send tens of thousands and the company has spent its AI budget on garbage... Sometimes it is the weapon.”
Severity scores miss the attack chain
More than one-quarter of the surveyed companies rejected the idea that isolated vulnerability scores provide enough information about practical risk. They argued that security teams should examine reachability, network position and the way several modest weaknesses can connect into a serious compromise.
Five companies described attack chains in which minor vulnerabilities, configuration errors and gaps in security controls combined to produce full administrative access. Amir Shavitt, Head of Research at Zafran, called the overlooked chain his “Moby Dick”: a sequence in which no individual weakness appears dangerous enough to trigger an urgent response.
The same concern extends to software outside the traditional enterprise perimeter. Nearly one-third of respondents identified open-source AI infrastructure, MCP bridges, GitHub and browser extensions as growing sources of risk. Mor Weinberger, a Security Researcher at Echo, said companies monitor endpoints and cloud environments while often failing to monitor the platforms where their code originates.
The survey’s practical warning is narrow and measurable: organizations have less time to move from discovery to remediation, while their AI systems add new permissions, dependencies and operating costs. A vulnerability is no longer assessed only by its severity on one machine; researchers increasingly judge it by the path it creates through the wider system.