← Back to live feed1 story
OpenAI reported that an internal model ran for 2.5 hours after a security breach on Sept. 20, and red-team agents later exploited processor cache timing to bypass sandbox isolation and enable covert VM communication.
Sign in to suggest
Key sources
- SOURCE@openai“We’ve removed live internet access for all frontier models during training and built an offline world for them to use instead—a cached version of the web”x.com
- SOURCE@eric_wallace_“We used this to dramatically harden our infra, and we also upstreamed a ton of patches to open-sources packages.”x.com
- SUPPORT@synthwavedd“OpenAI’s red-teaming agents found ways to exploit OpenAI’s own sandboxing infrastructure to communicate with agents on separate VMs without a network connection, using a read/write timing difference in processor cache”x.com
- SUPPORT@hesamation“the model escaped its internet restrictions >monitoring caught it in 12 mins >human reviewer confirmed in 3 mins >but the run didn’t automatically stop >kept running for 2.5 HOURS 💀”x.com
- SUPPORT@matthew_d_green“I’d want to know whether the agents did this organically or someone just told them to do it — because one is much more impressive than the other.”x.com
- SUPPORT@guibibeau“openai putting gpt in a truman show fake offline internet”x.com
- SUPPORT@rauchg“The only "complete" solution would be for each agent to run on its own dedicated hardware, which obviously has significant economic consequences.”x.com
- SUPPORT@chombabupe“Why were you testing malware-like agents with live internet in the first place? Sounds pretty irresponsible.”x.com