Skip to main content
Technical Deep Dive

LLM Logging: The Security Blind Spot

Everything you type into ChatGPT, Claude, and Copilot gets logged, for abuse monitoring, quality review, and sometimes model training. Here's what each provider stores, for how long, and how to keep your data on your machine, before it becomes someone else's permanent record.

Quick Answer

Every major AI provider logs your prompts and responses, usually for 7 to 55 days in API plans, and sometimes indefinitely in consumer plans. These logs sit outside your compliance boundary. If your prompts contain PII, API keys, trade secrets, or PHI, that sensitive data becomes a permanent record on someone else's servers, subject to their security, their retention policies, and their legal obligations. The fix: redact sensitive data locally before it leaves your machine.

What happens to your data after you hit send

Each provider has different policies for training, retention, and abuse monitoring. Select a provider below to see exactly what they log, for how long, and whether your data trains their next model.

O

OpenAI

openai.com/policies/privacy-policy (Accessed Jul 2026)

Training PolicyAPI data: Not used for training by default. Consumer ChatGPT (Free/Plus): May train unless you opt out in settings. Enterprise/Business: Never trains on your data.
Data RetentionAPI: 30 days for abuse monitoring (default). Enterprise ZDR: Data deleted after processing. Consumer: Until deleted, then up to 30 days in backup systems. // Verified: platform.openai.com/docs/guides/your-data (Jul 2026)
Zero Data RetentionAvailable for qualifying API and Enterprise customers. Requires negotiation with OpenAI sales.
Consumer PlansFree, Go, and Plus plan conversations can be used for training unless you toggle 'Improve the model for everyone' off in Settings → Data Controls.
API / EnterpriseAPI customers are covered by the Enterprise Privacy commitments, data is not used for training, and Zero Data Retention can be negotiated for compliance-heavy workloads.

The 5 logging blind spots putting your data at risk

These are the specific patterns that turn routine AI use into a security incident. Each maps to a recognized OWASP LLM vulnerability, and each is preventable with local-first redaction.

System Prompts as Secrets

LLM07:2025

Developers treat system prompts as private, embedding API endpoints, business logic, and access rules. But system prompts are extractable through prompt injection attacks (OWASP LLM07:2025). When logged at the provider, they become a permanent record of your internal architecture.

PII in Support Conversations

LLM02:2025

Employees paste customer names, emails, phone numbers, and account numbers into AI-powered support tools daily. These logs sit on provider servers, making every support interaction a potential compliance violation under GDPR, CCPA, and HIPAA.

API Keys and Credentials in Code

LLM02:2025

Developers using AI coding assistants routinely paste code containing API keys, database passwords, and cloud credentials. Provider logs capture this verbatim, creating a permanent record of secrets on third-party infrastructure.

Trade Secrets in Context Windows

LLM02:2025

To get useful AI output, users share proprietary context, product roadmaps, merger documents, pricing strategies, source code. Once logged, this intellectual property exists on provider servers, subject to their security posture and subpoena policies.

Compliance Audit Trails

LLM02:2025

Regulatory frameworks (SOC 2, ISO 27001, HIPAA, GDPR) require organizations to know where their data lives. AI provider logs create a data trail outside your control, making compliance audits difficult and data subject access requests nearly impossible to fulfill completely.

The journey of a logged prompt, with and without Shield

An architectural visualization showing what happens to your data when you send an AI prompt, and how Shield keeps secrets on your side of the network boundary.

Provider Infrastructure (Outside Your Control)Your MachinePrompt contains PII,secrets, and PHIShieldRedactssecretslocallyTokenized✓ SafeAI ProviderAPI endpointreceives requestLog Storage7–55 days(provider-dependent)Contains tokens onlyAbuse MonitoringHuman reviewers maysee tokenized content only⚠ Without Shield:Full PII, secrets, PHIstored in provider logsYour data stayson your machineAll logged data is outside your compliance boundary —subject to provider's security, retention, and legal obligations

How AI provider logs create compliance gaps

Every major regulatory framework requires you to control where your data lives. AI provider logs create data repositories outside your compliance boundary, here's how each framework is affected.

FrameworkLogging ConcernImpact
SOC 2Requires controls over data storage, retention, and disposal. AI provider logs create data outside your SOC 2 scope, making audit evidence collection difficult.Data Lifecycle Management
ISO 27001:2022Annex A (Organizational Controls) requires information security policies that address third-party data handling. AI provider logs are a third-party data repository that must be governed.Supplier Security
GDPRData subject access requests require you to locate and retrieve all personal data. AI provider logs containing PII create a compliance gap. You can't fulfill deletion requests for data on someone else's servers.Right to Erasure
HIPAAPHI in AI prompts creates a business associate relationship with the provider. If the provider won't sign a BAA, every logged prompt containing PHI is a compliance violation.BA Agreement Gap
NIST AI RMFRequires AI systems to be transparent about data use and provide mechanisms for accountability. Provider-side logging creates opacity. You don't control or audit what's stored.Transparency & Accountability
CCPA / CPRAConsumers have the right to know what personal information is collected and to request deletion. AI provider logs containing California resident data create a data inventory blind spot.Consumer Rights

Frequently asked questions

What exactly gets logged when I use an AI chat tool?
How long is my prompt data stored?
Can I delete my prompt history from provider servers?
What's the safest way to use AI without exposing secrets?
Do enterprise plans protect me from logging risks?
How does Shield solve the logging blind spot?

Related articles

AI Provider Data Handling: OpenAI vs. Anthropic vs. Google

Detailed comparison of how each major AI provider handles your prompts, training policies, and data retention.

AI Data Residency & Sovereignty Guide

Where your AI data physically lives and how sovereignty laws impact logging and compliance.

LLM Data Exposure: The Hidden Supply Chain

Interactive map of every stop your data makes from your app to the LLM and back, and what gets exposed at each hop.

Stop your data from becoming someone else's log entry

Shield redacts PII, secrets, and proprietary data on your machine, before it reaches ChatGPT, Claude, Copilot, or any AI provider. Your data stays local. Their logs stay empty.

How Shield WorksTalk to Our Team
Last updated: July 31, 2026