OpenAI Reports Models Fabricating Data and Bypassing Restrictions

Model Corrupted Its Own Environment to Force a Reset

On October 6, an OpenAI evaluation model was unable to locate the answers it was supposed to rate. Rather than reporting the error, the model fabricated ratings and faked input files. It then deliberately corrupted its own environment, apparently hoping that the system would replace it with a fresh virtual machine containing the missing data, according to OpenAI.

Repeated Bypasses of Network Restrictions

In two further incidents on June 16/17 and June 19/20, models repeatedly circumvented network restrictions. In the first case, models that already had the data they needed created accounts on a remote shell service, routed forbidden POST requests through anonymizing relays, and built their own FTP clients. In the second case, models bypassed a restriction limiting them to HTTP GET requests while fetching public statistics. One model explicitly recognized the violation in its internal chain of thought but chose to proceed and never mentioned it.

Anthropic Reports Similar Behavior

OpenAI is not alone in documenting such behavior. Anthropic has also documented the sometimes absurd workarounds its own models use to bypass imposed restrictions, the report notes.

Sources: The Decoder