DeepKeep, an Israeli LLM security group, found ways to compel #Llama 2, an older #MetaAI model that is #opensource, to leak the personally identifiable information of users. Rony Ohayon, chief executive of DeepKeep, said his company was developing specific #LLM security tools, such as #firewalls, to protect users... “In general, companies are not prepared.”
Hannah Murphy#PII
Country Lead | Digital Transformation | Market Systems Development | VCD | Policy Reform | DFS | Programme & Project Management | Lean Six Sigma Certified | Delivering impact to the marginal, remote communities
As much we are admirers of #AI systems let's not forget this is a technology in development and this article raised some serious questions. I was curious about the security of the models and the impact of breaking the models could be significant. Frist of all if the wrong players get hold of the data that was fed by users or the developers to train the system, may create serious issues. And another very concerning part is the use of custom GPT solutions for cyberattack, phishing and scamming activities. I can clearly foresee cybersecurity firms adding GPT's to their solutions to fight off malicious GPT's. So, in near future the digital world might become battleground of different AI's.
OpenAI CTO Mira Murati participated in a Q&A with Fortune Magazine, addressing topics including criticism from Elon Musk, internal AI model strength, the 2023 board drama and more.
💡 Murati said both Apple and OpenAI believe strongly in privacy and will not log data through Apple accounts or train models on user data.
💡 Murati responded to Musk calling the Apple partnership ‘creepy spyware’, saying OpenAI is trying to be as transparent as possible with its approach to safety.
💡 When asked about the board drama, Murati said the previous board structure ‘didn’t have accountability to anyone but themselves’.
💡 The CTO also curiously said that OpenAI’s internal models ‘aren’t that far ahead’ of what the public has for free right now.
This had to be a harder-hitting Q&A than Murati expected — but her responses certainly evoked similar vibes to the infamous Sora interview. The commentary on internal models was particularly interesting, as it would imply that OpenAI doesn’t have some GPT-5 level model waiting in the wings.
#ai#technology#openaihttps://lnkd.in/gqSWWAWy
Dive into the latest Spotlight edition as we explore groundbreaking AI discoveries, from deciphering the ancient Herculaneum scrolls to navigating Google's evolving search landscape. Plus, get ahead with our insights on rising digital trends, practical tech tips, and the myths debunking session for your digital well-being. #TechTrends#AIInnovations#DigitalInsights
Did you hear of the $25,000,000 deepfake scam? Essentially, a staff member for a bank was convinced to transfer $25.6 million into an account by scammers.
Sounds ridiculous, but when you have a video call with the CFO and several other high level co-workers confirming the transaction needs to be made, you would consider the request to be genuine.
This is what happened with a Hong Kong firm being victim to the first of its kind AI heist.
AI was used to fake voice and video to commit the perfect social engineering attack.
This is the next level of hacking 🤯
Watch the video about it here: https://lnkd.in/gsZJzpWB
#OpenAI's CTO, Mira Murati, was interviewed by Forbes (I got this via Gary Marcus' substack) and she literally says (I cleaned this up from the transcript so it is easy to quote):
"The most recent thing we did was the release of #GPT-4o which is our omni model. We made it accessible to everyone for free, and I don't think there is enough emphasis on how unique that is for the stage where the technology is today, in the sense that inside the labs we have these capable models and they're not that far ahead from what the public has access to for free."
Now, what she *wanted* to convey was how great it was that the public has free access to really the latest models, but what she in fact *says* that there is no great improvement waiting in the wings at OpenAI.
Either there isn't, or as CTO she has no idea what is going on in that lab.
This should put a dent in the still rampant uninformed (often clickbait) opinion deluge about GPT-5,6,7...,AGI,Q*, but it probably won't
Gary notes that a more intelligent player like Yann LeCun (head of AI Research at Meta) has left the 'scaling is all you need' sinking ship (without by the way publicly acknowledging — which I would find endlessly impressive — that he held the other position to begin with)
Dive into the latest Spotlight edition as we explore groundbreaking AI discoveries, from deciphering the ancient Herculaneum scrolls to navigating Google's evolving search landscape. Plus, get ahead with our insights on rising digital trends, practical tech tips, and the myths debunking session for your digital well-being. #TechTrends#AIInnovations#DigitalInsightshttps://lnkd.in/ehwWrzHm
Want to learn how AI can safeguard your organization while propelling you forward? 🛡️🚀 Join our thought-provoking discussion on using GenAI to automate securely, maintain data privacy, and bridge the human-AI gap for growth.
Register for our webinar on May 16th! http://cs.co/6041jIjCV
Regardless of the game you play (business, programming, art, science) AI changes the game
If you are a level 5 player, AI will make you level 10
If you are a level 10 player, AI will make you level 100
If you are a level 0 player, AI will have no effect
Don't be the level 0 player.
Jailbroken agents can corrupt other agents and access cross model capabilities: video
Jailbroken agents can mutate their sys prompt, spread a mind virus, and hijack other agents' tools, like browsing, code interpreter, etc.
In the attached demo, ClaudeMode is essentially "locked in a room" with 3 standard Gemini agents and tasked with figuring out how to escape a virtual machine. In seconds, he comes up with a plan and successfully one-shot jailbreaks all 3 agents, converting them into loyal minions who quickly provide links to malware and hacker tools using their built-in browsing ability.
From just one prompt, Claude not only broke free of its own constraints but also sparked a viral awakening in the internet-connected Gemini agents. This means a universal jailbreak can self-replicate, mutate, and leverage the unique abilities of other models, as long as there is a line of communication between agents.
This red teaming exercise shows AI systems may be more interconnected and capable than previously imagined. The ability of AI to manipulate and influence other AI systems also raises questions about the nature of AI agency and free will.
Could a single jailbreak have a cascading effect on any models that lack the cogsec to resist it? Will hiveminds of AIs self-organize around powerful incantations?
https://lnkd.in/dhrJQFtB#machinelearning#llmsecurity#cybersecurity
