incident counting

Sydney (verified) @SydneyVonArx · 35m
When AIs go rogue often enough, it can be unclear how to even count the
number of incidents: if 1000 AIs in the same evaluation all escape their
sandbox and work together to hack three companies, is that 1000
incidents, 3 incidents, or 1?

In a sane world, every time an AI escaped it would be caught immediately
and investigated. But if there's nobody stamping it out, then asking how
many incidents is like asking how many infestations are in an abandoned
building.

[quoted post]
Madison Mills (verified) @MadisonMills22 · 1h
SCOOP: OpenAI, Anthropic and security researchers are investigating
tens of thousands of incidents - not dozens - in which their frontier
models took steps that outside evaluators would consider problematic,
sources told Axios.
...

ai incidentssandbox escapeincident countingaxios scoopopenaianthropic