Palo Alto Networks expects advanced offensive artificial intelligence capabilities to become widely accessible within four to six months, based on how quickly previous generations of AI models were replicated or matched by open-source alternatives, the US cybersecurity company’s chief information officer said.
The industry may have only a narrow window before sophisticated offensive AI capabilities become widely accessible through open-source models, dramatically lowering the barrier for cyberattacks, Meerah Rajavel told ET in an interview. The company has already seen a surge in customer concerns following public discussions around Anthropic’s Project Glasswing and the emergence of increasingly capable AI systems, she said.
Palo Alto Networks is one of 12 core primary partners in Project Glasswing, an initiative involving early access to Anthropic’s Claude Mythos Preview model, and has also evaluated newer generations of large language models from other providers.
“Last month has been intense,” she said. “Customers that were planning security upgrades over 12 months are now compressing those timelines into weeks or months.”
According to Rajavel, who was in Bengaluru last week as part of an India visit, Palo Alto Networks has built an internal testing harness and conducted large-scale red teaming exercises with the models by simulating real-world cyberattacks. “We uncovered vulnerabilities and attack paths that would normally have taken us a year to find,” she said.
The models demonstrated three major capabilities that alarmed cybersecurity teams: identifying vulnerabilities at scale, chaining together seemingly low-risk flaws into sophisticated attack paths and generating technical code to launch attacks in parallel.
“Humans typically prioritise critical and high vulnerabilities and often ignore low ones because of the volume involved,” Rajavel said. “But these models are very good at chaining multiple vulnerabilities together and creating novel attack paths that humans may not think about.”
She warned that the threat could intensify once comparable capabilities become available through open-source AI systems. “The frontier models can still impose some guardrails. Open source cannot,” she said.
The company’s recommendations to enterprises centre on four broad steps: rapidly assessing infrastructure vulnerabilities, implementing continuous attack surface management, adopting zero-trust architectures and building AI-assisted autonomous security operation centres.
Rajavel said traditional patching cycles are no longer adequate in the AI era.
“Most enterprises patch monthly or weekly. But when attacks can be weaponised in minutes, that’s no longer sufficient,” she said. “You have to fight AI with AI.”
Palo Alto Networks chief executive Nikesh Arora told ET in February that while companies typically take between four and 14 days to detect intrusions, bad actors could now infiltrate and exfiltrate data in as little as eight minutes.
Palo Alto Networks internally operates in what Rajavel described as a “zero-day patching mode”, where a majority of security patches are deployed within a day or two of release.
She also highlighted the growing risk to operational technology (OT) systems used in critical infrastructure sectors such as banking, telecom, utilities and manufacturing.
“The OT components are very hard to patch. Sometimes patching doesn’t even exist as a concept,” she said. “You have to think about shielding them, air-gapping them, or blocking exploitability paths altogether.”
For Indian enterprises, the stakes are particularly high given the country’s rapidly expanding digital economy, large developer ecosystem and growing base of global capability centres, she said.
India has emerged as one of Palo Alto Networks’ fastest-growing markets over the last five to seven years, with strong adoption across banking, IT services, manufacturing and pharma sectors.
Rajavel said cybersecurity discussions have now moved into boardrooms.
“CEOs and board members are asking, 'What should I worry about? What are the right questions to ask?',” she said. “This is no longer viewed as crying wolf. Companies understand this is real.”
She cautioned against treating cybersecurity as an afterthought in AI adoption strategies.
“You cannot treat security like seasoning,” Rajavel said. “It has to be like oil in cooking, built in from the beginning.”
“The ease and the speed are completely changing the risk factor,” Rajavel said. “It’s not that these attacks were impossible earlier. We could do them. But now, with AI, the pace and scale have changed the game.”
