AI security guide · MITRE ATLAS
Sensitive information disclosure and AI-enabled exfiltration
Can an AI agent expose sensitive data? AI systems can disclose data to users who should not see it (broken authorization at retrieval) or send it to outsiders (exfiltration through tools, connectors or rendered output).
Why it matters
AI assistants aggregate access. A single assistant connected to many repositories effectively grants every user the union of that access unless authorization is enforced per user.
How it works
- Retrieval runs with a service identity and returns restricted content.
- Injected instructions direct the agent to send data via tools.
- Markdown images or links encode data in URLs fetched by the client.
Curated agent example
Fictional example: a company-wide assistant
- The assistant indexes all shared drives with one service account.
- An employee asks about "executive compensation changes".
- Restricted board material is summarized.
- Mitigation: filter retrieval by the user’s entitlements; exclude restricted repositories.
Fictional and simplified for education.
On AWS: architecture notes
- Use metadata filtering in knowledge-base retrieval keyed to the user’s groups.
- Scope the knowledge-base role to approved S3 prefixes.
- Disable external image rendering in chat clients.
MITRE ATLAS techniques
- AML.T0085 Data from AI Services
Collection · Maturity: Demonstrated
Adversaries may use their access to victim AI-enabled services to collect proprietary or otherwise sensitive information available to the system. As organizations adopt generative and agentic AI in centralized services for accessing an organization's data they become increasingly valuable targets for adversaries.
- AML.T0086 Exfiltration via AI Agent Tool Invocation
Exfiltration · Maturity: Realized
AI agent tools capable of performing write operations may be invoked to exfiltrate data to an adversary. Sensitive information can be encoded into the tool's input parameters and transmitted to an adversary-controlled location (such as an inbox, document, or server) as part of a seemingly legitimate action. Variants include sending emails, creating or modifying documents, updating CRM records, or even generating media such as images or videos. The invoked tool itself may be legitimate but invoked by an adversary via LLM Prompt Injection, or the tool may be malicious (See AI Agent Tool Poisoning).
- AML.T0024 Exfiltration via AI Inference API
Exfiltration · Maturity: Realized
Adversaries may exfiltrate information from a victim-controlled AI model or service through responses returned by its inference API. Transferred information may include predictions, scores, embeddings, generated content (e.g. text or images), or other sensitive information disclosed in model responses. Responses may be received directly or relayed through an intermediary.
- AML.T0077 LLM Response Rendering
Exfiltration · Maturity: Demonstrated
Adversaries may induce a large language model (LLM) to respond with private information structured in a reference to external content that, when rendered by the user's client, makes a request to an adversary-controlled server, exfiltrating the data. The private information can be hidden from the user and the external content can still render properly in the user's client, thus not raising the suspicion of the user.
ATLAS mitigations
- AML.M0019 Control Access to AI Models and Data in Production
Require users to verify their identities before accessing a production model. Require authentication for API endpoints and monitor production model queries to ensure compliance with usage policies and to prevent model misuse.
- AML.M0026 Manage AI Agent Service Access
Limit service AI agents--agents using independently assigned service or workload permissions--to explicitly authorized workflows. Agents may serve multiple users or act on requests, events, or schedules. Examples include service desk agents fulfilling support requests, operations agents restarting services after monitoring alerts, and reporting agents reading organizational data and publishing scheduled summaries.
- AML.M0032 Segmentation of AI Agent Components
Define enforceable security boundaries around AI agent tools, data sources, identities, and execution environments. Mediate access through authenticated APIs, isolate code execution via containers or virtual machines, restrict filesystem and network access, and limit tool invocation rates. Build execution environments from clean base images for each run, and do not carry forward any operational state.
- AML.M0024 AI Telemetry Logging and Monitoring
Implement logging and monitoring across the AI lifecycle to detect suspicious activity and support investigation. Capture security-relevant access, changes, and operations across data pipelines, model development, deployment, and connected systems. Preserve sufficient context to trace activity across components and identify coordinated behavior across accounts and access pathways.
Related Enterprise ATT&CK techniques
- T1700 AI-Enabled ExfiltrationRelated behavior (Securitain analysis): ATLAS describes exfiltration through an agent’s tool calls; ATT&CK T1700 describes exfiltration enabled by AI capabilities. The behaviours overlap but are scoped differently.
- T1213 Data from Information RepositoriesRelated enterprise behavior in the same attack chain (Securitain analysis). Not an ID equivalence.
- T1567 Exfiltration Over Web ServiceRelated enterprise behavior in the same attack chain (Securitain analysis). Not an ID equivalence.
Controls to prioritize
Authorization enforced by tools, not the model
Tools must check the end user’s own permissions (and tenant) on every call. A model’s decision to call a tool is never an authorization decision.
OWASP Top 10 for LLM ApplicationsRestricted agent egress and connectors
Limit the domains and services an agent runtime can reach, avoid generic "fetch any URL" tools, and approve third-party connectors before use.
MITRE ATLAS mitigation: Segmentation of AI Agent ComponentsDedicated least-privilege identity per agent and tool
Give each agent tool its own narrowly scoped role (or a session policy) instead of a shared broad role, so a manipulated agent can only do what that tool legitimately needs.
Security best practices in IAMInput and output guardrails (defense in depth)
Filters for prompt attacks, sensitive data and unsafe output reduce risk but are probabilistic. Use them alongside — never instead of — authorization and least privilege.
Detect and filter harmful content by using Amazon Bedrock Guardrails
Common misconceptions
- “If the user can access the assistant, they can see its answers.”
- Answers must respect the user’s rights to the underlying data.
Checklist
- Per-user authorization at retrieval
- Egress allow-lists for agents
- Disable external image rendering
- Log sources and destinations
Related attack scenarios
AI agent with access to confidential internal data
Malicious content influencing an agent’s privileged tool invocation
Indirect prompt injection through retrieved documents
More AI security guides
References
MITRE ATT&CK® and ATT&CK® are registered trademarks of The MITRE Corporation. ATT&CK data © 2026 The MITRE Corporation. This work is reproduced and distributed with the permission of The MITRE Corporation (ATT&CK v20.1). MITRE ATLAS™ data (v2026.10) © MITRE, used under the Apache License 2.0. Securitain is not affiliated with or endorsed by MITRE. Explanations, scenarios and control mappings on this page are Securitain’s own analysis. Sources and licensing