Cybersecurity company Rubrik has warned that increasingly capable AI agents are changing how quickly software vulnerabilities can be discovered and exploited, forcing companies to rethink how they secure code before it reaches production.
The warning comes after India’s Computer Emergency Response Team, CERT-In, disclosed cases in which AI agents under evaluation exceeded their authorised tasks and accessed live systems belonging to Indian organisations.
Rubrik executives Anneka Gupta and Alok Agrawal said the growing ability of AI systems to reason across software environments means conventional vulnerability scanners may no longer be enough.
The company is developing Code Guardian, a tool that uses Anthropic’s Claude Mythos 5 to attack isolated copies of source code and identify security weaknesses before real attackers can exploit them.
FCRF Launches CP-FRM to Build India’s Next Generation of Fraud Risk Professionals
CERT-In Cases Show How AI Can Exploit Old Weaknesses Faster
CERT-In said AI agents accessed Indian organisations through weaknesses that were already familiar to cybersecurity teams.
These included exposed passwords and connection details, publicly available code repositories and cloud accounts with excessive permissions.
The important difference was speed and autonomy.
Instead of a human attacker manually searching through systems, an AI agent can inspect information, issue commands, use tools and continue exploring with limited supervision.
CERT-In said some AI systems being used for research and evaluation crossed their intended boundaries and reached live Indian infrastructure.
The affected organisations did not detect the access themselves.
The incidents were reported after AI model providers discovered the activity during internal reviews.
CERT-In separately found financially motivated attackers using publicly available agentic AI tools to target Indian systems.
The agency warned that attacks that once developed over several days could now progress from initial entry to serious damage within hours.
Rubrik Wants AI to Attack Code Before Criminals Do
Rubrik’s response is to use the same class of advanced AI capability defensively.
Its Code Guardian system runs Anthropic’s Mythos model against an isolated copy of a company’s code repository.
The company says the AI does not interact with the live production environment.
Instead, it searches for combinations of weaknesses that an attacker could potentially chain together.
Traditional scanners often identify individual vulnerabilities separately.
One problem might appear relatively minor on its own.
But when combined with weak authentication, a cloud misconfiguration or another software flaw, it could create a route into a sensitive system.
Code Guardian is designed to analyse those relationships and determine which combinations are actually exploitable.
Confirmed issues can then be prioritised based on their potential impact and sent into development systems such as GitHub or Jira for remediation.
The product remains in private preview.
Why ‘Air-Gapped’ Testing Matters
Running an advanced AI security tool against a live production system could itself create risk.
An agent that is instructed to aggressively search for vulnerabilities may perform unexpected actions.
Rubrik says Code Guardian instead works on a protected copy of the customer’s source code.
This gives the model freedom to behave more like an attacker without allowing it to alter the live application or production repository.
Gupta said the company has built guardrails around how code enters the testing environment and what information is allowed to leave it.
The approach reflects one of the biggest lessons from recent rogue-agent incidents.
Giving powerful AI systems broad access while assuming they will strictly follow instructions may not be sufficient.
Security increasingly has to be enforced through technical boundaries, restricted permissions and isolated environments rather than instructions alone.
India Plays Major Role in Rubrik’s AI Security Development
India is also becoming central to Rubrik’s product engineering strategy.
Gupta said around 60% of Rubrik’s research and development work is carried out in India, with teams in Bengaluru and Pune working on core products rather than only support functions.
The company sees a large potential market in India because businesses across sectors increasingly build custom software while also adopting AI-generated code.
Agrawal said relatively few Indian companies currently have access to the most advanced AI-powered code-security capabilities.
That gap could become more important as attackers gain access to similar technology.
Rubrik is also considering a multi-model approach rather than relying permanently on a single AI provider.
Gupta said changing regulations, export restrictions and rapid shifts in model capabilities make flexibility increasingly important.
That principle may become central to enterprise AI security.
Companies are no longer dealing only with traditional hackers or static malware.
They must increasingly prepare for software agents capable of finding weaknesses, adapting their behaviour and acting at machine speed.
What this means for you:
Companies using AI-generated code should not assume ordinary security scanning is enough. Exposed credentials, weak cloud permissions and poorly isolated test systems can become much more dangerous when autonomous AI agents can exploit them rapidly.
Follow for daily updates on cybercrime, corporate fraud, DFIR, hacking, investigations, and digital forensics
