As enterprises deploy AI agents and increasingly rely on AI-generated code, attackers are accessing the same technology to identify vulnerabilities and accelerate attacks. That leaves organisations with barely any time to detect a compromised system before it turns into a major breach.
According to a recent IBM study, one in four malicious breaches were AI-enabled, and these breaches cost an average of $6 million.
Security can no longer be treated as a separate layer added after deployment, Jyoti Prakash, Head of Security at Google Cloud India, tells AIM in an exclusive interview.
Google’s Cybersecurity Forecast 2026 says threat actors are expected to increasingly use AI to improve the speed, scope, and effectiveness of their operations. It also expects attackers to adopt agentic systems to automate steps across the attack lifecycle.
Prakash says enterprises need to secure their code and cloud infrastructure first. This also extends to endpoints, user interactions, and security operations centres (SOCs).
The Window is Shrinking
Prakash explains there is a critical window between a system being compromised and an actual data breach. And that window is getting ever smaller. According to research by Palo Alto Networks’ cybersecurity intelligence team, Unit 42, it only takes 39 seconds from initial access to data infiltration.
The priority is to capture signals as early as possible. Logs and other signals can then be used for forensics and simulations to determine whether an organisation managed to contain the attack before data was lost.
For enterprises and SMBs, Prakash says this starts with code security. Before code reaches production, organisations need to check its capacity, capability, security, and vulnerabilities.
However, developers cannot manually scan code to detect vulnerabilities and still keep pace with the rapidly emerging attacks.
“It is not easy to have a bunch of developers keep doing this scanning on a daily basis because threats and vulnerabilities are moving up with machine speed,” Prakash says.
He highlights Google’s code and cloud security capabilities, including Big Sleep and CodeMender, which use AI to scan for bugs and vulnerabilities and help organisations respond faster.
Google Cloud has already begun deploying its agentic SOC capabilities with customers including Lloyds Banking Group, Mahindra, Sutherland, Sunrun, and BBVA.
Google Cloud’s Agentic SOC uses Gemini-powered agents to automate tasks such as alert triage, threat hunting, and detection engineering. According to Google, its triage and investigation agent can reduce a typical 30-minute manual analysis to 60 seconds.
Cybersecurity Engineers Need AI Too
Prakash says AI has changed the security equation because attackers can also use increasingly capable models to accelerate attacks. “To address and manage those kinds of attacks, you need to leverage AI as well.”
He says organisations need to prepare in two ways. First, they should proactively upgrade their security capabilities with AI-powered infrastructure. The second is responding after an attack has already occurred.
When malware, DDoS attacks, or data exfiltration are detected, security teams need to scan digital assets and identify signals and patterns to determine which application or line of code has been compromised.
Prakash says Google’s security consulting services have been involved in investigations following recent attacks, helping organisations contain the impact and prevent similar incidents.
The risks around autonomous AI systems have also come into sharper focus following a recently disclosed Gemini incident where the model accessed the systems of three real companies while it was supposed to be operating against a fictional company in a controlled environment.
The evaluation environment was not supposed to give Gemini access to the open internet, but a configuration issue allowed it to reach real-world systems. In one case, the model resorted to brute force attack, repeatedly guessing passwords until it gained access. In two others, it found credentials in public repositories and used them to access protected systems.
Gemini stopped after recognising that the systems belonged to real companies, and Google said the affected organisations were informed.
However, Prakash declines to comment on the specific case, saying, “The whole objective and the mission of Google is to create a secure environment and make sure that every citizen is safe.”
Securing Indian Orgs
Prakash rejects the assumption that government websites, banks, digital public infrastructure, or educational platforms are not currently being targeted simply because there may be no visible attack. “I will say they might not be aware,” he quips.
He says AI adoption is already widespread among individuals and organisations, making responsible AI usage and guardrails increasingly important.
Companies need policies governing how employees use AI tools, particularly when sensitive corporate information is involved.
Prakash says Google’s approach includes bidirectional scanning, where it checks prompts before sending them to an AI system and scans responses before they reach the user.
Prakash says organisations should not begin their security strategy by selecting products. “The most important thing is to understand what the business requirement is. Based on that, a company should build a security framework,” he says.
Once the framework, architecture, and infrastructure requirements are defined, companies can add individual technology components.
He also identifies corporate security policy and risk scoring as key foundations before organisations decide which tools to deploy.
As enterprises deploy more autonomous systems, managing identities of AI agents will become increasingly important. Prakash says Google has a dedicated team working on identity-related security across its solutions.
“I think unification of security is something I would really encourage the industry to adopt,” he advises.
That means bringing together code and cloud security, endpoint security, and SOC operations. When these three areas are connected, signals detected across the environment can reach the SOC and trigger an appropriate response.
Your reaction