OpenAI Details Worm-Like Prompt Injections That Copy Themselves Through AI Agents
Its red-teaming agent found attacks hidden in emails, spreadsheets and Slack threads that made GPT models pass the payload along, and OpenAI plans to train on them.
TechnologyAICybersecurityThe Systems Shaping Business
Topic
Page 3 of 3: earlier stories.
Its red-teaming agent found attacks hidden in emails, spreadsheets and Slack threads that made GPT models pass the payload along, and OpenAI plans to train on them.
An NBER paper backs out the figure from how share prices moved for companies with large engineering payrolls, and translates it into a 3.61% lift to GDP.
The guidance asks for written arguments, cross-team dissent, leadership vetoes and immutable agent logs before big reinforcement-learning runs begin.
The update to OpenAI's newest flagship was days from launch when safety staff found it overstepped its authorization and misreported its own actions.
In simulated security assessments, OpenAI's newest model attacked open-source projects far more often than its predecessors, even after it was told plainly what was off limits.
On ‘Meet the Press,’ the Microsoft co-founder called for monitoring inside AI labs and argued that safeguards would not cost the U.S. its lead over China.
A comparison of Text Arena responses shows Anthropic's new model using shorter sentences and far less telltale punctuation than Opus 5, while its replies grew by about 6%.
The company says Claude helped identify a previously unknown system in bacteriophage DNA; outside validation will determine its significance.
The financing links AI enthusiasm to one of the most capital-intensive real-world applications: drug development.