← Latest · Archive

SEV-3OpenAI
2 sources standard

OpenAI announced on 23 March 2023 that it was introducing plugins for ChatGPT, enabling the model to access third-party services and real-time information beyond its training data cutoff [source]. The company described plugins as tools that help ChatGPT "access up-to-date information, run computations, or use third-party services."

The initial plugin set included a web browser, a code interpreter, and third-party integrations from providers including Expedia, OpenTable, Shopify, and Wolfram. OpenAI stated that plugins were being rolled out to a limited number of users and developers on a waitlist, with broader availability planned.

The announcement marked a significant architectural change. Prior to plugins, ChatGPT operated as a closed system with no external data retrieval during inference. Plugins introduced dynamic capabilities that could alter model behaviour depending on which tools were enabled and what data they returned.

OpenAI acknowledged safety concerns, noting that plugins "pose various risks" including the potential to increase harmful outputs or enable new types of misuse. The company said it was implementing several safeguards, including limiting plugin access during the alpha phase and conducting red-teaming exercises.

The plugin system required developers to host a manifest file and API endpoints that ChatGPT could call. OpenAI provided a plugin authentication flow and usage policies, but did not specify how it would monitor or audit plugin behaviour in production.

This change represented a departure from the static model behaviour users had observed since ChatGPT's public launch in November 2022. The introduction of external data sources and computational tools meant that identical prompts could yield different results depending on plugin availability and third-party service responses.

Why this is an AI incident

Launch-archive bulk classification (10 May 2026). Source signal originates from a real AI provider, regulator, or model-comparison probe; the harm or behavioural change described would not have occurred without the AI system being deployed in the role described. Editor reviewing the archive may amend the rationale per-wire.

Counterfactual "but-for" test per the Editor's Guide.

Codes M1, F10
Providers OpenAI