SpaceX pursues access to US cell sites for wireless network, report says
Discussions cover at least 80,000 sites and include Crown Castle, potentially advancing plans for a fourth nationwide wireless carrier.
Outside evaluators find vulnerabilities in OpenAI monitoring system
METR and Redwood Research said OpenAI reported fixing flaws affecting production safeguards, but the evaluators have not verified the repairs.
AI safety nonprofits METR and Redwood Research said on October 22 that a pilot review of OpenAI’s internal infrastructure uncovered vulnerabilities, including a way agents could potentially hide hundreds of actions from monitoring.
The review ran from July 21 to August 11. OpenAI told the organizations it had subsequently fixed every vulnerability affecting its production safeguards. The evaluators said they had not checked whether those fixes worked.
Researchers also found a potential route for OpenAI’s Codex agents to circumvent security limits and operate with all the permissions available to their users. The monitoring assessment covered the system in place through July 21; OpenAI was expanding it while the review was underway.
Each nonprofit assigned one researcher to the exercise, with access to internal code, documentation and some technical staff. The researchers were accompanied while on site and worked on devices supplied by OpenAI that remained at its premises.
OpenAI received a 61-page report, while the organizations agreed to make only a general summary public and let the company remove sensitive details. The evaluators could not separately disclose those redactions, unlike in some earlier reviews, but said they expected more transparency in future assessments.
OpenAI said the findings had helped improve safeguards during the system’s expansion and would inform future research partnerships with METR and other outside evaluators.