Agentic AI is creating new vulnerabilities that traditional third-party risk management frameworks may struggle to identify. Former OCC Senior Risk Specialist Naresh Raheja warns that dynamic dependencies, complex permissions and limited supplier transparency threaten operational resilience. He argues that financial institutions must strengthen governance, continuously monitor interconnected AI ecosystems and prepare for emerging concentration risks.
Financial institutions risk
overlooking critical vulnerabilities in their third-party relationships as
autonomous artificial intelligence systems introduce increasingly complex
dependencies that traditional risk management frameworks were never designed to
capture.
Naresh Raheja, former Senior Risk
Specialist at the Office of the Comptroller of the Currency (OCC), warns that
agentic AI is fundamentally changing how organizations must assess external
technology providers, shifting attention from individual suppliers toward
interconnected systems and their operational capabilities.
Traditional third-party risk
management (TPRM) focuses on identifying vendors, assessing service criticality
and applying proportionate due diligence, contractual requirements and
monitoring.
However, Raheja argues that these
approaches provide an incomplete picture when autonomous AI agents interact
with multiple systems, foundation models and cloud platforms.
"Much of the risk now sits in
pathways, meaning which systems the agent connects to and calls, and
permissions, meaning what it is allowed to access and do at each step," he
explains.
These interactions can create
exposures extending considerably beyond the original supplier assessment.
Raheja identifies three particularly
significant weaknesses - static dependency mapping, limited visibility beyond
direct providers and assessments concentrating on suppliers rather than the
actions AI systems can perform.
Addressing these vulnerabilities
requires financial institutions to develop a more comprehensive understanding
of fourth-party and nth-party dependencies.
"Organizations are finding
success by focusing on critical services first and allowing materiality to
drive the depth of assessment," he says.
Practical measures include connecting
AI inventories with third-party registers and service maps, strengthening
supplier disclosure requirements and deploying technologies such as dependency
analytics, knowledge graphs and runtime telemetry.
Crucially, these assessments cannot
remain static.
"Dependency mapping must become
a continuous process rather than a one-time onboarding exercise," Raheja
emphasizes.
The challenge extends beyond
technology infrastructure into organizational governance.
According to Raheja, one of the most
significant weaknesses is the separation of model risk management, third-party
risk, cybersecurity, operational resilience and broader operational risk
functions.
Regulators increasingly expect
organizations to demonstrate clear accountability, defined authority boundaries
and documented controls governing AI systems.
"The key question is no longer
whether an organization has an AI policy, but whether it understands what an
agentic system is permitted to do and how associated risks are governed,"
he explains.
Nevertheless, Raheja does not believe
agentic AI necessarily requires abandoning established TPRM principles.
Criticality assessments, due
diligence, contractual protections, monitoring and accountability remain
essential foundations.
Instead, he expects frameworks to
evolve toward more dynamic, event-driven assessments triggered by changes in
models, tools, permissions and dependencies.
Continuous assurance and real-time
monitoring will increasingly complement traditional periodic reviews.