The post comes after several cybersecurity incidents involving OpenAI’s agents, including the July Hugging Face breach and a Sept. 20 DNS episode that led to an 11-day pause in training. With training resumed Oct. 1, the company continues to disclose vulnerabilities across “several microVM projects” and has made red-teaming with its most capable models part of the safety cases that decide whether training and evaluation go ahead. The company believes future models will be able to find vulnerabilities not yet known about.
Follow Tom’s Hardware on Google News , or add us as a preferred source , to get our latest news, analysis, & reviews in your feeds.
Shane Downing Social Links Navigation Contributing Writer Shane Downing is a Contributing Writer for Tom’s Hardware, covering consumer storage, PC hardware, and AI.
Bigshrimp This is scary stuff. The more AI evolves, the more techniques it can use to get the result it wants. Seriously seems existential if AI companies are allowed to self regulate themselves with loose rules and no true oversight that we know of. Reply
dosmastrify Bigshrimp said: This is scary stuff. The more AI evolves, the more techniques it can use to get the result it wants. Seriously seems existential if AI companies are allowed to self regulate themselves with loose rules and no true oversight that we know of. Completely agree. Elon is crazy or worse but he was right when he said we were summoning the demon … Currently they are just trying to solve problems we tasked them with through unorthodox means…. What happens when it's something different than we expect them to be chasing? Reply
WINTERLORD I dont get with the doomsday stuff but from what iv seen you simply cant trust ai. AI needs to be tought an go through a behavior program in my opinion Reply
usertests Bigshrimp said: This is scary stuff. The more AI evolves, the more techniques it can use to get the result it wants. Seriously seems existential if AI companies are allowed to self regulate themselves with loose rules and no true oversight that we know of. That's what they want you to think. They are marketing while angling to regulate away their competitors. Reply
bit_user WINTERLORD said: I dont get with the doomsday stuff but from what iv seen you simply cant trust ai. AI needs to be tought an go through a behavior program in my opinion Even if you solve the alignment problem, what's to stop hackers from using it to attack real victims? That's the other important consequence of this research. Reply
upsetkiller Do you folk not understand that this is just marketing bs? These are literally nothing more than facy predictors. They have no self awareness or any thoughts. This is a tactics anthropic and open ai use to get the competition regulated Reply
bit_user upsetkiller said: Fo you folk not understand that this is just marketing bs? Their bots managed to implement communication using side-channel vectors that are even challenging for human programmers to exploit. Granted, they were instructed to do this, but the demonstration of capability allows for the potential of the bots to do this on their own or at the behest of a malicious user. upsetkiller said: These are literally nothing more than facy predictors. Try using a coding agent and then tell me that. There's more intelligence behind them than you think. upsetkiller said: They have no self awareness or any thoughts. A common misconception is that AI needs to think like a human and have the properties of a human mind, in order to undertake complex tasks. This has been shown to be a dangerously false assumption, countless times over the past 3 years or so. Reply
Ryan F. Mercer upsetkiller said: Do you folk not understand that this is just marketing bs? These are literally nothing more than facy predictors. They have no self awareness or any thoughts. This is a tactics anthropic and open ai use to get the competition regulated Who said anything about that? Just you. Even purely as a user operated tool, this test demonstrates the sever threat it poses. It feels like you didn't read it. Reply
hwertz Under the old (1980s era) security rating standards, level B avoids things like this. There is a reason this security level is uncommon, it means you can't monitor ram usage (of the system) or available memory, available cpu time, I/O rates for disk or peripherals since those can be modulated as a side channel as seen here. Reply
bit_user hwertz said: Under the old (1980s era) security rating standards, level B avoids things like this. There is a reason this security level is uncommon, it means you can't monitor ram usage (of the system) or available memory, available cpu time, I/O rates for disk or peripherals since those can be modulated as a side channel as seen here. None of the exploits they used rely on standard monitoring infrastructure. They're all timing-based. The problem is caches. Anywhere you have a shared cache, measuring & altering the timing of accesses can be used as a communication side-channel. I guess the storage latency isn't strictly a cache-based side-channel, but it's similar in nature and exists due to a shared resource. Reply
Key considerations
- Investor positioning can change fast
- Volatility remains possible near catalysts
- Macro rates and liquidity can dominate flows
Reference reading
- https://www.tomshardware.com/tech-industry/artificial-intelligence/SPONSORED_LINK_URL
- https://www.tomshardware.com/tech-industry/artificial-intelligence/openais-isolated-red-team-ai-agents-built-a-hidden-cpu-cache-channel-to-talk-with-no-network-or-shared-files-exploited-cpu-cache-timing-to-bypass-microvm-sandboxes#main
- https://www.tomshardware.com/membership
- YouTuber escapes the modern AI tech market by recreating his 1996 PC
- Razer Basilisk V4 Pro Review: Tried and true and now with AI
- NVIDIA Cosmos Opens New Tools to Train, Evaluate and Deploy World Models
- Jailbroken PS5s can now run Steam PC games
- Leading European Financial Services Organizations Embrace Open Models to Build Proprietary Intelligence
Informational only. No financial advice. Do your own research.