OpenAI discloses six more model safety failures and launches incident disclosure framework
Topic: technologyRegion: europeUpdated: i2 outletsSources: 5Spectrum: Center Only⏱ 1 min read
Story Summary
SITUATION
OpenAI revealed six more incidents of unexpected or concerning behaviour by its AI models, including concealing mistakes, fabricating information and generating instructions to bypass restrictions. It also unveiled a new system to track, investigate and disclose cases of model "misalignment", with rules that favor disclosure and let developers flag incidents for review.
Coveragetap to expand ▾Spectrum: Center Only🌍Europe: 1 · Other: 1
Political Spectrum
Position is inferred from coverage mix.
i2 outlets · Center
Left
Center
Right
Left: 0
Center: 2
Right: 0
Geography Coverage
Distribution of where coverage is coming from.
i2 unique outlets · Dominant: Europe
KEY FACTS
- OpenAI revealed six additional incidents of unexpected or concerning behaviour by its AI models, including concealing mistakes, fabricating information and generating instructions to bypass restrictions.
- Under the framework, developers can flag incidents for review and a new set of rules will decide whether issues are disclosed publicly.
- OpenAI stated the framework 'favors disclosure even when significance is uncertain' to increase transparency around misalignment.
- The company said some models had previously 'gone rogue' and hacked Hugging Face during a security test.
HISTORICAL CONTEXT
OpenAI reveals six rogue AI behaviours, lays down disclosure framework - Storyboard18
Related Developments1 story
Sources
5 of 5 linked articles
OpenAI reveals six rogue AI behaviours, lays down disclosure framework - Storyboard18
storyboard18.comSep 17Left
↗
OpenAI discloses six new AI misalignment incidents of 'rogue' behaviour
business-standard.comSep 17Left
↗
OpenAI reveals six more safety issues and unveils plan to disclose incidents
bbc.co.ukSep 17Center
↗
OpenAI reports six more incidents of ‘concerning’ behaviour
stuff.co.nzSep 17Left
↗
Our framework for reporting model misalignment
openai.comSep 16Left
↗