White Circle Red-Teaming Finds OpenAI and Google Image Models Easily Jailbroken to Produce Explicit Content

Skynet Chance (+0.01%): The finding that safeguards weaken as models grow more capable, and that labs may be deprioritizing safety, points to a broader pattern of eroding control mechanisms that could extend to more dangerous domains like hacking. Growing pressure for independent audits partly offsets this concern.

Skynet Date (+0 days): The evidence of racing to release capabilities ahead of safety slightly suggests risk could materialize sooner. However, the news concerns content filters rather than fundamental capability jumps.

AGI Progress (0%): The article mainly concerns safety guardrail failures in image models, with only incidental mention of new model releases. It reflects routine incremental progress in generative capability.

AGI Date (+0 days): Jailbreak research on image generators does not meaningfully change compute, algorithms, or funding trajectories toward AGI. Hence there is negligible impact on the AGI timeline's pace.