Agentic AI... The Newest Attack Surface, and the Least Secured.
Jim Leone
11/19/20253 min read
Artificial intelligence has evolved faster in the past 24 months than most platforms have evolved in 20 years. But the biggest shift isn’t the models, it’s the agents. Agentic AI systems can already browse the web, write and execute code, deploy infrastructure, read and respond to emails, analyze logs, perform triage, interact with APIs, and take actions without waiting for a human.
In other words... We are entering a world where AI doesn’t just “assist”, it operates.
And that creates a completely new cybersecurity problem... AI agents now have privileges, identities, access, and decision-making power traditionally reserved for administrators.
I believe the question every SOC leader should be asking is--> Who is securing the machines that act like humans? Because right now, it's clear to me that the answer is... almost no one.
The Emerging Risks of Agentic AI That I Believe Security Teams May Be Underestimating...
Prompt Injection Becomes RCE for AI
When an agent reads emails, documents, logs, or webpages, a hidden instruction can hijack it...
Exfiltrate data
Alter firewall rules
Modify SOAR runbooks
Create new agents
Disable monitoring
We spent decades fighting code injection. Now we must fight instruction injection.
Uh Oh! AI-to-AI Lateral Movement
Organizations are chaining multiple agents together...
Agent A queries data →
Agent B triages →
Agent C remediates
Compromise one agent, and you compromise the entire chain. This would be machine-speed lateral movement.
Where's The Audit Trail For Accountability?
When an agent closes a ticket, blocks an IP, rotates a key, or provisions cloud resources, did it...
Log the action?
Explain why?
Use proper approvals?
Today, from what I've seen, most agent systems have zero auditability.
Data Poisoning That Leads to Operational Misfires
Feed an agent manipulated...
Threat intel
API responses
Logs
Metrics
Email content
…and it will take incorrect actions with absolute confidence.
This may be the first time in history where data integrity failures can cause automated operational actions.
Over-Privileged Agent Credentials
To “make the demo work,” early adopters grant agents...
Administrator API keys
Cross-account access
Broad IAM roles
Write privileges to production
Let's not forget that just one compromised agent = full compromise.
Multi-Agent Decision Loops
Agents now...
Write tasks
Assign tasks
Approve tasks
Execute tasks
These loops multiply risk. One manipulated decision spirals across the entire system.
This is the AI version of a supply-chain attack.
Today's Agents Lack Human Judgment
Humans hesitate before dangerous actions. Agents don’t. If the logic is wrong, an AI can...
Terminate production workloads
Delete data
Overwrite configurations
Trigger auto-remediation loops
And it can do it in milliseconds!
No (Accepted) Frameworks or Controls Exist Yet
Zero Trust was built for...
users
devices
networks
workloads
But not autonomous AI. We don’t have...
Agent identity
Agent least privilege
Agent accountability
Agent kill switches
Agent provenance
Agent certificates
Agent drift detection
I feel this may be the largest security gap of the early AI era.
Emerging Frameworks & Tools (A New Industry Is Forming)
The good news... security minds have started to address this gap. Here are the most notable early frameworks, research efforts, and toolsets.
BeyondTrust - Agentic AI Security Platform
One of the first commercial solutions explicitly targeting...
AI agent identity
Least privilege for AI
Zero standing access
Just-in-time permissions for autonomous systems
OWASP - State of Agentic AI Security & Governance (1.0)
A major step forward. Defines...
Threats
Governance models
Lifecycle controls
Agent oversight standards
This should become required reading for SOC and security architects.
AAGATE - Agentic AI Governance Assurance & Trust Engine (Research)
Defines a runtime governance layer for autonomous agents...
Policy engine
Behavior validation
Agent certificates
Identity delegation
“Trust transmission” between agents
I understand that this is academic now, but it foreshadows what enterprise controls will look like.
SAGA - Security Architecture for Governing Agentic Systems (Research)
Addresses...
Agent identity
Permissioning
Cryptographic signing
Delegation chains
Verifiable action records
We can think of this as “Zero Trust for Agentic AI.”
Microsoft, NIST, and ISO Early Work
All three are beginning to define...
AI safety baselines
Autonomous system standards
Inference integrity requirements
System-of-agents risk definitions
Nothing enterprise-ready yet, but coming fast.
What I Believe Agentic AI Security Must Look Like (The Future State)
Based on everything above, here’s the emerging consensus of what “secure agentic AI” will require...
Agent Identity (AID)
Agents must authenticate like services, certificates, tokens, rotation.
Least-Privilege Agents
Not “read/write all.” Scoped roles. Narrow tasks. Guardrails.
"Human-in-the-Loop" for High-Risk Actions
Agents should recommend... humans authorize.
Explainability Before Execution
Agents must justify actions... “Here is what I intend to do and why.”
Immutable Audit Logs
Every decision. Every action. Every input. Logged, timestamped, verifiable.
Agent Kill Switches
If an agent behaves unexpectedly --> immediate shutdown.
Continuous Behavior Monitoring
Anomaly detection for agents, not just networks.
This Matters! And Why I Feel We Must Act Now...
Agentic AI will soon -->
write code
deploy infrastructure
tune firewalls
respond to alerts
interact with customers
handle billing
manage accounts
This is the next digital workforce. But unlike humans, agents don’t sleep, don’t hesitate, don’t second-guess, and don’t question bad instructions. We must secure them before they scale, not after. Organizations that adopt agentic AI without governance are building the next generation of breaches into their foundations.
The IP HighWay
Stay updated with the latest IT security news.
info@iphwy.com
© 2025. IPHwy LLC. All rights reserved.



