The Amazon MadPot findings mark a meaningful shift in the threat landscape. Attackers are no longer probing AI systems from the outside. They are operating inside the same agentic architectures that enterprise teams are actively building, using autonomous agents to harvest credentials at scale, persist across knowledge stores, and move laterally through AI infrastructure that most security teams have not yet classified as attack surface. If your organisation is running LLM-backed applications or agent pipelines in production, the threat model has changed and the security controls need to catch up.
Companion piece to our broader work on agentic attack surfaces. See AI Agent Security Risks: Attack Surface Guide for a detailed breakdown of identity gaps, MCP exposure, and control plane risks revealed at Black Hat.
LLM Gateway Credentials Are Now a Primary Target
LLM gateways sit at the centre of most enterprise AI stacks. They route prompts, manage rate limits, handle retries, and often hold API keys for multiple upstream model providers in a single configuration layer. That concentration of credentials makes them a high-value target, and the attack surface is broader than most teams recognise.
Keys are frequently embedded in environment variables, CI/CD pipeline configs, or application builder templates where they are accessible to anyone with read access to the deployment environment. Automated agents scanning public repositories and misconfigured cloud storage buckets can harvest these credentials without any human operator in the loop. The MadPot campaign demonstrated that this harvesting is now industrialised, not opportunistic.
The commercial implication is direct. A stolen gateway key does not just expose model costs. It exposes every downstream system the gateway is authorised to reach, including retrieval stores, vector databases, and any tool-calling integrations the agent is configured to use.
Agentic Frameworks Introduce Orchestration-Layer Risk
Open-source agentic frameworks have matured quickly. They give engineering teams the ability to build multi-step, tool-using agents with relatively little infrastructure overhead. That speed is operationally useful, but it introduces a category of risk that sits above the model layer and below the application layer, in the orchestration logic itself.
Trust Boundaries in Multi-Agent Systems
When one agent delegates a task to another, the receiving agent typically inherits the calling agent's permissions unless the framework explicitly enforces a trust boundary. Most open-source frameworks do not enforce this by default. An attacker who can inject a malicious instruction into one agent's context can propagate that instruction through a chain of agents, each acting with the same authority as the original caller.
Tool-Calling as an Escalation Vector
Tool-calling agents interact with external APIs, file systems, and databases as part of their normal operation. If an agent's tool permissions are not scoped to the minimum required for a given task, a compromised agent becomes a pivot point into adjacent systems. This is the agentic equivalent of a misconfigured IAM role, and it carries the same consequence: lateral movement at machine speed.
Knowledge Graph Persistence Creates a Durable Foothold
Many enterprise agent deployments now include long-term memory components: vector stores, knowledge graphs, or structured retrieval layers that persist context across sessions. These stores are operationally valuable precisely because they retain information over time. That same property makes them attractive for persistence.
An attacker who can write to a knowledge graph, either through a compromised agent or a poisoned retrieval pipeline, can embed instructions or false context that influences future agent behaviour without any further access to the system. The payload survives session boundaries, key rotations, and even redeployments if the knowledge store is not wiped and revalidated. Security teams familiar with database persistence techniques will recognise the pattern; the difference is that the payload here operates through natural language rather than SQL.
The Controls That Close the Gap
Treating AI infrastructure as a first-class security asset means applying the same discipline that mature organisations apply to cloud IAM. That means several concrete engineering decisions.
Credential Governance for Model APIs
Gateway API keys should be managed through a secrets management system with automatic rotation, not stored in environment variables or application configs. Each service or agent that needs model access should hold its own scoped credential, not a shared key. Audit logs for key usage should feed into the same SIEM pipeline as cloud credential events.
Agent Identity and Least-Privilege Entitlement
Every agent in a production pipeline needs a persistent, verifiable identity. That identity should carry an entitlement model that specifies exactly which tools, APIs, and data stores the agent is authorised to access for a given task. Permissions should not be inherited transitively across agent-to-agent calls without an explicit grant.
Knowledge Store Integrity Controls
Write access to long-term memory components should be restricted to authenticated, audited processes. Retrieval pipelines should validate the provenance of stored content before surfacing it to an agent. Periodic integrity checks against a known-good snapshot are a practical baseline for detecting tampering.
Raising AI Infrastructure to the Level of Cloud IAM
The underlying argument here is about classification, not just controls. Most enterprise security programmes have mature frameworks for cloud IAM, network perimeter, and application secrets. AI infrastructure, including LLM gateways, agent orchestration layers, and knowledge stores, has been treated as developer tooling sitting adjacent to those frameworks rather than inside them.
The MadPot campaign makes that classification untenable. Attackers are already treating these components as primary targets. The security function needs to catch up, which means bringing AI infrastructure into existing asset inventories, threat models, and incident response playbooks rather than treating it as a separate domain managed exclusively by the AI team.
The technical controls are available. The gap is organisational: deciding that LLM gateway credentials, agent permissions, and knowledge store integrity belong in the same governance conversation as S3 bucket policies and IAM role boundaries. That decision is a CTO-level call, and the window for making it proactively is narrowing.
Where Vector Labs Fits
We design and build production agent infrastructure with security architecture embedded from the start, not retrofitted after deployment. In our agent identity analysis, we set out the engineering architecture most teams are missing: non-human identity governance, least-privilege entitlement models, and audit trail design for agentic systems in production. If you are building agent infrastructure and need a security architecture review before it reaches production, contact us at vector-labs.ai/contacts.
FAQs
Gateway credentials typically authorise access to multiple upstream model providers and, in many architectures, to the retrieval stores and tool integrations connected to those models. A single compromised gateway key can therefore expose a much broader set of systems than a conventional application API key scoped to a single service. Attackers using automated agents can also exploit stolen keys at a rate and scale that makes the cost exposure significant before the incident is detected.
In a tool-calling agent, a malicious instruction injected into the agent's context can direct it to exfiltrate environment variables, call an external endpoint with sensitive data, or write poisoned content to a shared memory store. The agent executes these instructions as part of its normal task loop, which means the action can occur without any direct attacker access to the underlying infrastructure. The risk scales with the breadth of the agent's tool permissions.
A poisoned knowledge graph can alter agent behaviour across every future session that retrieves from that store, without the attacker maintaining any ongoing access to the system. The injected content can include false context, manipulated instructions, or data designed to bias the agent's outputs in a specific direction. Because the payload persists through session boundaries and may survive key rotations, detection requires integrity monitoring of the store itself, not just access log review.
Cloud IAM governance principles map well to agent infrastructure: least-privilege access, scoped credentials per service identity, audit logging for all privileged actions, and regular entitlement reviews. NIST SP 800-207 (Zero Trust Architecture) provides a useful conceptual baseline for enforcing trust boundaries between agents. The main adaptation required is extending these frameworks to cover non-human identities, including agents, rather than treating them as extensions of the human user who deployed them.
The most practical starting point is an asset inventory that explicitly includes LLM gateways, agent orchestration services, vector stores, and knowledge graphs alongside cloud resources and application services. Once these components appear in the asset register, they can be assigned owners, classified by sensitivity, and subjected to the same vulnerability and configuration review processes that apply to the rest of the stack. Governance cannot follow assets that the security programme does not know exist.

