@saraislet@infosec.exchange2026-09-30 05:36:44
"We’re also sharing a preview of specialist dots with their own identity for access management, IT-provisioned hardware, and support for deep integrations with a company’s systems of record to allow them to take on well-defined responsibilities inside organizations."
IAM: dots
IT: dots
D&R: dots
someone who is good at the security please help me threat model this. my company is getting hacked
#GPT #OpenAI #AgenticOverlords
@comex@mas.to2026-09-24 19:15:34
It's kind of crazy that both GPT and Claude really love to manipulate files by writing Python scripts that do text search-and-replace. First of all, why isn't this a dedicated tool? Second, the generated scripts usually don't check that the search-and-replace was successful. Are they just good enough at regurgitating text that they never make mistakes? What if the file was modified in the mean time?
@Techmeme@techhub.social2026-10-22 18:36:32
Microsoft says its GPT-5-powered Diagnostic Orchestrator uses agents that mimic panels of experts to diagnose conditions up to 4x more accurately than doctors (Robert Hart/The Verge)
https://www.theverge.com/ai-artificial-int
@veit@mastodon.social2026-08-31 12:47:02
GPT 5.6-Cyber escapes from a VM on multiple occasions. A standard VM with frequent updates will therefore no longer be sufficient in future. Only virtualisation solutions with a minimal attack surface, such as Firecracker, are likely to still offer a degree of security:
#GPT #AI #VM #ITSecurity
@comex@mas.to2026-10-18 07:08:07
Wow. Finally got around to setting GPT (6.1 Sol) at Mario Maker 2's collision detection (something I had previously partially reverse engineered by hand), and it just breathes out the kind of understanding that took me hours and hours of staring at Binary Ninja, carefully labeling things. I hate it. I liked staring at Binary Ninja.
@pre@boing.world2026-10-13 23:30:16
They tell me that Open AI have solved four hundred previously open maths problems this week.
The news media, and even the mathematicians, seem quite credulous about it.
No way to know how many of those will really stand up in the end. Lots of proofs turn out not to be really proofs coz they have a mistake, and AI is quite good at discovering mistakes as well as now apparently discovering maths.
Some of them are counter-examples that prove the conjecture is false and those are gonna be good at last. Machines can check it.
Some have other types of mechanical machine-checkable proof, which is interesting in that we can all believe the checker-machine: coz it's deterministic and understandable.
We can't trust the machine that isn't deterministic or understandable as much though.
GPT says these things are true. Now maths people have to check, because OpenAI staff aren't expert mathematicians either, mostly.
Some of the 400 proofs will fall, I'm sure.
How many tokens will be spent before machines find the flaws that are surely still there? How many tokens until you decide to stop?
But if a machine can prove a thing to itself, but not explain it to any person, is it really doing maths? Even if it's correct more often than people are? Or is doing maths more about making meat understand a thing?
🤔
I'm sure I can't tell if the proofs are true or not. No point me reading them. Mostly won't even understand the question. The boundary of my meager meat-mind's maths-understanding was around barely-degree level.
Can the academic discipline continue as one which is devoted mostly to trying to understand the output of a machine writing about maths, which is mostly right but nobody can really tell without an entire academic discipline trying to check.
In other words, mathematicians now also face a spam problem familiar to many github project leads.
#ai #maths