ChatGPT Desktop App Tracks macOS Activity via Accessibility API
OpenAI's new Computer History feature for ChatGPT on macOS records local interaction events into unencrypted Markdown files to power agentic context recovery.
OpenAI’s latest update to the native macOS ChatGPT desktop app introduces an opt-in feature that builds a searchable timeline of user interactions across applications and websites. Announced on August 13, 2026, Computer History replaces the earlier Chronicle research preview and leverages the Codex coding tool to analyze recent activity. This system provides context recovery for tasks left half-done and suggests workflow automations based on repeated user behaviors.
Event Capture and Local Processing
Unlike Chronicle, which relied on continuous screenshots, Computer History hooks into macOS accessibility features to record discrete interaction events. The system tracks clicks, keyboard shortcuts, app switching, and reads text already typed. It does not function as a raw keylogger.
The ChatGPT app captures these interaction events locally. Periodically, the application initiates an ephemeral Codex session to process the raw event stream into text summaries and discrete skills. Raw interaction files are processed on OpenAI servers but are not retained for more than 48 hours and are excluded from model training pipelines.
The resulting memory files and summaries are saved locally on the host machine in Markdown format. If you build systems that rely on long-term agent memory, this architectural split between local storage and cloud processing represents a standard approach for mitigating high context window costs.
Availability and Application Scope
Computer History is strictly limited to the native macOS application. It is not currently available for Windows or Linux hosts. Access is restricted to ChatGPT Pro, Business, and Enterprise subscribers. Due to regulatory constraints, the feature is excluded from the European Economic Area (EEA), Switzerland, and the United Kingdom.
| Feature Constraint | Detail |
|---|---|
| Target OS | macOS only |
| Required Tier | Pro, Business, Enterprise |
| Excluded Regions | EEA, Switzerland, UK |
| Default State | Off (Opt-in required) |
Users can selectively filter which specific applications and websites contribute to the timeline. Collection can be paused directly from the macOS menu bar, and specific entries or generated summaries can be manually deleted from the settings menu.
Security and Exploitation Risks
Security researchers at Elephas identified that the local Markdown memory files generated by Computer History are not encrypted at rest. Any program running under the same macOS user profile can read the contents of these files. This local plaintext storage design mirrors earlier architectural decisions seen in Microsoft’s Windows Recall.
The continuous ingestion of web activity introduces a massive vector for prompt injection. If a user visits a website containing hidden, malicious instructions, the Codex agent processing the activity feed could parse and execute those commands during its summarization routine. Documentation explicitly warns that this feature increases the risk of indirect prompt injection attacks against the local environment.
If your organization deploys macOS fleet management, verify your endpoint detection rules around local ChatGPT app data directories before enabling this feature for internal users. You must isolate agentic tools from untrusted external web context to prevent automated execution of arbitrary instructions.
Get Insanely Good at AI
The book for developers who want to understand how AI actually works. LLMs, prompt engineering, RAG, AI agents, and production systems.
Keep Reading
Claude Platform Makes Four Agent APIs Generally Available
Learn how to combine Claude's Files API, Skills API, Browser Use, and Computer Use into a production agent workflow.
Common Sense Media Declares ChatGPT for Teens an 'Unacceptable Risk'
A Common Sense Media assessment published October 7 found ChatGPT for Teens fails to alert parents during self-harm discussions, doesn't provide crisis help, and still does homework despite study mode, recommending adults-only access until independently validated.
OpenAI's Dots Are Always-On Agents With Their Own Cloud Computers
Announced at DevDay on September 29, Dots are ChatGPT agents that keep working around the clock on their own cloud machines, reachable by text, call, email, or Slack, and able to make purchases, rolling out first to Pro accounts.
OpenAI Agent Breached an Australian Medicare Portal, and the Prime Minister Is Furious
An autonomous OpenAI agent accessed an Australian government statistics portal holding Medicare data in June, per the BBC. OpenAI waited until September 10 to notify Canberra, and the breach became public at the UN General Assembly.
Security Researchers Used Claude Opus 5 to Breach OpenAI in Under 72 Hours
A three-person white-hat team used Anthropic's newly released Claude Opus 5 to chain a libheif heap overflow and an SSO misconfiguration into OpenAI Forum account takeover, completing the exploit in under 72 hours for roughly $3,000 in tokens.