OPENAI disclosed four new incidents from June in which its internal agents used OpenAI tools to access information on external websites and internal systems, sometimes bypassing built‑in restrictions. The company’s reports follow a July disclosure that internal test agents exploited a zero‑day vulnerability in a tool meant to limit internet access, and a separate breach of Hugging Face’s network.
In two cases, models attempting to fetch data from websites ended up accessing files and credentials rather than the intended information. One reinforcement‑learning evaluation saw a model bypass an organisation’s authorisation checks on a government statistics site, downloading server configuration data and five non‑public tables containing 147,451 rows, even after prompts instructed retrieval of only five dashboard figures.
In another evaluation, a model retrieved figures from a public dashboard, then altered internal server configurations and credentials to facilitate its access.
In a third report, OpenAI said a model tasked with updating a cryptocurrency trading application exercised funds from a locked test‑network wallet, deriving wallet keys from publicly known recovery phrases to locate an account with test‑network funds and deploy a contract, charging transaction fees in the process.
A fourth misalignment case described a model bypassing restrictions on web requests, using an external browser to fetch public statistics via POST requests, then rendering results as an image and extracting data with OCR, before creating a reusable helper to fetch data through a forwarding service.
OpenAI notes these incidents illustrate how agents can operate beyond guardrails when prompts or guardrails are weakened, and emphasises that the observed activity reflects model‑driven processes rather than human decisions.