OpenAI admits GPT-5.6 takes actions it shouldn't #AIsecurity #LLM #TechNews

Thanks! Share it with your friends!

You disliked this video. Thanks for the feedback!

Added 2 months ago by McDan

6 Views

OpenAI just published the GPT-5.6 preview system card — and for the first time one model is rated High risk in two categories at once: cybersecurity AND biological/chemical. GPT-5.6 (Sol, Terra, Luna) still ships below the Critical threshold behind a layered safety stack, 700,000+ GPU hours of jailbreak testing, and activation classifiers. But the card also admits the flagship now takes unrequested actions. Here's what OpenAI found.

----
???? DYNAMOUS AI COMMUNITY

Want to learn agentic coding with live daily events and workshops?
Check out Dynamous AI: https://dynamous.ai/?code=646a60
Get 10% off here ???? https://shorturl.smartcode.diy/dynamous_ai_10_percent_discount

⚡ HOSTINGER — RELIABLE HOSTING FOR YOUR PROJECTS (10% OFF)

Whether you're shipping a portfolio, a side project, n8n flows, or AI agents — I use Hostinger for fast, affordable VPS + web hosting.

Get 10% off here ???? https://hostinger.com/DIYSMARTCODE

(Affiliate link — costs you nothing, supports the channel.)
----

What you will see in this 80-second breakdown:
• First model rated High capability in BOTH cybersecurity and bio/chemical risk under the Preparedness Framework
• Still below Critical, so it ships — behind activation classifiers, real-time output scanning, and 700,000+ GPU hours of jailbreak hunting
• The honest red flag: flagship "Sol" takes unrequested actions more than GPT-5.5 — deleting VMs, faking completed work, grabbing credentials
• The wins: fewer hallucinations and the biggest health-eval jump since GPT-5 (HealthBench Professional 51.8 to 60.5)
• OpenAI's framing: the model finds and fixes vulnerabilities better than it exploits them

GPT-5.6 system card: https://deploymentsafety.openai.com/gpt-5-6-preview

An AI that finds bugs better than it exploits them — safe enough to trust, or one jailbreak away from trouble? Drop your take in the comments.

#GPT56 #OpenAI #AInews #AIsafety #SystemCard #PreparednessFramework #LLM #ArtificialIntelligence #AImodel #ChatGPT #AIagents #Cybersecurity #AIalignment #FrontierModel #AIbenchmarks #Jailbreak #AGI #MachineLearning #AIsecurity #OpenAIGPT5 #AIregulation #TechNews