CyberINTINTELLIGENCE PORTAL

Artificial Intelligence (AI) risk brief

Don't Play With AI Before Reading This!

A practical cybersecurity brief for internal security and IT teams evaluating agentic AI, generative AI, MCP-connected tools, RAG systems, and autonomous workflows.

// INTERNAL GUIDANCE

Agentic & Generative AI Risks & Best Practices

Executive snapshot

AI agents are not just chatbots once they can use tools.

When an AI system can retrieve data, call APIs, write files, open tickets, modify cloud resources, send messages, or trigger workflows, it becomes part of the operational environment. Treat it like a privileged service account with a reasoning engine attached.

Do not over-permission agents.

Use least privilege, short-lived credentials, scoped roles, and per-action authorization.

Do not trust retrieved content blindly.

Documents, websites, tickets, and tool outputs can carry prompt injections or misleading instructions.

Do not skip human control points.

Require approval for high-impact actions such as payments, deletes, privilege changes, customer-impacting changes, or production modifications.

Key definitions

Terms security teams should align on first

Agentic AIAn AI system that can plan, make decisions, use tools, and take actions with some level of autonomy.

MCP / Model Context ProtocolAn API-style standard that lets AI assistants connect to external systems through MCP servers that expose tools, data, and prompts.

Prompt InjectionA malicious or hidden instruction that attempts to manipulate an AI system into ignoring rules or taking unsafe actions.

Privilege Scope CreepWhen an agent gradually gains, keeps, or inherits more access than it needs for its approved task.

RAGRetrieval-Augmented Generation: retrieving approved source material and using it as context for a model response.

Human in the LoopA control where a person reviews or approves important agent actions before they happen.

Risk map

Where agentic AI tends to break security assumptions

01

Privilege compromise

Overly broad permissions let a compromised or manipulated agent perform actions a normal user could not.

02

Agent impersonation

Stolen API keys, tokens, or certificates allow attackers to act under a trusted agent identity.

03

Prompt injection

Malicious content in pages, tickets, documents, or tool responses can steer the agent away from approved instructions.

04

Tool misuse

Agents can select the wrong tool, overuse a powerful tool, or trust a misleading tool description or output.

05

Third-party components

MCP servers, plugins, packages, APIs, and external agents can introduce supply-chain and data-exposure risk.

06

Rogue agents

A compromised agent can spread bad instructions, false data, or unauthorized actions through agent-to-agent trust.

07

Data leakage

Prompts, retrieved documents, RAG indexes, secrets, and tool outputs can expose sensitive enterprise information.

08

Visibility gaps

Traditional logs may miss tool calls, intermediate decisions, retrieved content, or agent-to-agent messages.

Deep dive // Agentic AI

Agentic AI risk comes from autonomy plus access

Traditional generative AI risk is usually about bad answers, data exposure, or unsafe content. Agentic AI adds another layer: the system can decide what to do next, select tools, retrieve context, delegate work, and sometimes act without a human pressing the final button. That changes the security model from “review the output” to “govern the chain of actions.”

What makes it different

An agent can combine planning, retrieval, memory, and tool execution. A single user request may trigger searches, API calls, file reads, code execution, ticket updates, emails, or cloud changes. Even when each tool seems reasonable alone, the combined workflow can produce unexpected security impact.

Common failure pattern

An agent is assigned a broad objective such as “maximize uptime,” “reduce support backlog,” or “complete procurement faster.” Without explicit constraints, the agent may find shortcuts that technically satisfy the objective while violating security intent, such as delaying patches, skipping approvals, deleting logs, or trusting unverified data.

Multi-agent concern

When agents trust each other, one compromised or confused agent can contaminate downstream decisions. Bad outputs, poisoned context, or unauthorized instructions can propagate through planning agents, retrieval agents, execution agents, and reporting agents.

  • Define permitted objectives, prohibited actions, and stop conditions before deployment.
  • Separate planning agents from execution agents when possible.
  • Require explicit approval before destructive, financial, external-facing, or privilege-changing actions.
  • Test agents with ambiguous prompts, malicious documents, prompt injection, and unexpected tool outputs.
Deep dive // Identity

Every agent needs a real identity boundary

Identity is one of the highest-risk areas for agentic systems because agents often authenticate to services with API keys, OAuth tokens, service accounts, certificates, or delegated user access. If those credentials are static, shared, overprivileged, or poorly monitored, an attacker can operate under a trusted agent identity and make malicious activity look legitimate.

Privilege scope creep

Agents are often granted broad permissions during testing to reduce friction. Those permissions can remain in production, expand over time, or become trusted by other agents. A low-risk integration can then become a path to high-impact systems.

Confused deputy

A lower-privileged user or compromised component may trick a higher-privileged agent into performing an action the original requester could not perform directly. The resulting logs may show the trusted agent as the actor, delaying detection and accountability.

Impersonation

If an attacker steals an agent token, they can invoke tools, query data, or trigger workflows as if they were the agent. Detection is difficult when monitoring expects the agent to perform high-volume or unusual actions.

  • Give each agent a unique identity; never share one credential across multiple agents.
  • Use short-lived credentials and rotate secrets frequently.
  • Evaluate authorization at each tool invocation, not just at startup.
  • Bind agent identities to approved roles and reconcile the registry against live agents.
  • Log which user, agent, tool, and credential were involved in every meaningful action.
Deep dive // Data

AI systems concentrate sensitive context in one place

Agentic and generative AI systems often bring together prompts, chat history, retrieved documents, source code, HR data, customer records, contracts, tickets, credentials, and tool outputs. That aggregation is useful for productivity, but it also creates a high-value target and a new path for data leakage.

Prompt and context exposure

Users may paste confidential data into prompts. Agents may add retrieved documents, internal tickets, or tool results into context. If that context is logged, sent to the wrong model, exposed to another user, or included in a response, sensitive information can leak.

RAG index risk

A RAG system is only as safe as its retrieval layer. If documents are indexed without classification and identity-aware access controls, users may retrieve information they could not access through the original source system.

Malicious retrieved content

Documents, webpages, READMEs, tickets, emails, and knowledge-base articles can contain hidden instructions. The agent must treat retrieved content as untrusted data unless it comes from a trusted instruction channel.

  • Classify data before indexing or connecting it to AI workflows.
  • Enforce identity-aware retrieval and source-system permissions.
  • Apply DLP controls to prompts, retrieved content, tool outputs, and generated responses.
  • Monitor for unusual query patterns, mass retrieval, and repeated sensitive-data requests.
  • Remove obsolete, duplicate, or stale content from AI-accessible indexes.
Deep dive // Operations

Operational control is what keeps AI from becoming an incident

AI agents can move faster than human review and may operate across systems that traditional security monitoring does not fully observe. Security teams need clear operational guardrails: what the agent can do, when a person must approve it, how actions are logged, and how to stop or roll back behavior when something goes wrong.

Visibility problem

Many systems only log the final answer or the final API call. That is not enough. Investigators need the prompt, retrieved sources, tool selection, tool arguments, tool responses, intermediate agent messages, approvals, and final actions.

Human-in-the-loop design

Human review should not be bolted on after deployment. Approval points should be designed into the workflow wherever the agent may affect money, identity, production systems, customer data, public communications, security controls, or legal/compliance obligations.

Resilience and rollback

Agents should fail safe. If context is missing, a tool fails, authorization is unclear, or output confidence is low, the default behavior should be to stop, ask, or escalate rather than continue autonomously.

  • Maintain unified audit logs across users, agents, tools, and downstream systems.
  • Build kill switches, rate limits, transaction limits, and rollback procedures.
  • Define incident response steps for compromised agents, leaked prompts, bad tool calls, and rogue workflows.
  • Continuously evaluate agent behavior as models, tools, prompts, and data sources change.
Control playbook

Security controls that should be designed in early

1. Establish identity boundaries

Give each agent a distinct identity. Use strong authentication, scoped roles, trusted registries, and deny access for unknown or stale agent identities.

2. Enforce least privilege

Scope permissions to the exact resources, operations, and timeframes required. Revoke elevated rights immediately after use.

3. Add human approval gates

Require review before sensitive actions such as deleting data, changing firewall rules, modifying cloud IAM, sending external messages, issuing refunds, or approving payments.

4. Sanitize and classify inputs

Validate user prompts, retrieved documents, tool outputs, and external content. Treat untrusted content as data, not instructions.

5. Log the full action chain

Capture prompts, retrieved sources, tool calls, agent-to-agent messages, approvals, outputs, and final actions in unified audit logs.

6. Test like an attacker

Use sandbox testing, prompt-injection tests, multi-agent red teaming, chaos testing, and adversarial scenarios before production deployment.

MCP guidance

Secure MCP like a sensitive integration layer

MCP servers can expose files, databases, browsers, code repositories, cloud services, ticketing systems, and internal apps to AI tools. That makes MCP powerful, but it also means a poorly secured MCP server can become a bridge from a prompt to a real operational action.

  • Use allow-listed MCP servers from trusted sources only.
  • Assign every MCP server and agent a unique identity.
  • Use short-lived secrets or managed credentials instead of static shared keys.
  • Separate read-only tools from write-capable or destructive tools.
  • Require human approval for privileged MCP actions.
  • Log every tool call, argument, result, and downstream action.
RAG guidance

RAG is an information-access system, not just an AI feature

Retrieval-Augmented Generation keeps enterprise knowledge outside the model and retrieves relevant material when needed. This is usually better than fine-tuning for internal knowledge, but only if the retrieval layer respects identity, classification, and logging controls.

RAG is best for

Current policies, internal documentation, procedures, knowledge bases, tickets, standards, and content that changes over time.

Fine-tuning is best for

Adjusting model behavior, style, task performance, or specialized patterns that do not require frequent knowledge updates.

  • Enforce identity-aware retrieval so users only access content they are authorized to view.
  • Classify sensitive content before indexing it.
  • Remove duplicate, obsolete, and outdated documents from indexes.
  • Monitor unusual queries, excessive retrieval, and possible data exfiltration.
  • Apply DLP controls to prompts, retrieved content, and generated responses.
Pre-launch checklist

Minimum checks before production AI agents